Initial release: RyWorld VLN stage1 discrete step15000 ckpt + eval pkg

.gitattributes

@@ -1,35 +1,3 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.tar.gz filter=lfs diff=lfs merge=lfs -text

EVAL_SUMMARY.md

@@ -0,0 +1,70 @@
+# RyWorld VLNVerse Eval — Full coarse/val_unseen 835 ep
+
+## Final metrics (canonical: dedupe-by-trajectory_id across all 835 ep)
+
+| Metric | Value |
+|--------|-------|
+| **SR (success rate)** | **51.14%** |
+| **SPL** | **49.22%** |
+| **OSR (oracle success rate)** | **64.79%** |
+| **NE (navigation error, lower better)** | **3.727 m** |
+| **TL (trajectory length)** | **6.121 m** |
+| **nDTW (path alignment)** | **0.9445** |
+
+## Run config
+- Checkpoint: `ckpt_step0015000_final.pt` (stage1 discrete, InternVL3.5-1B + StopHead + ProprioProjector)
+- `stop_threshold` = **0.95** (selected after 4-point sweep 0.88/0.92/0.95/0.97 on 30-ep smoke)
+- Sharded 8-way: 7×104 + 1×107 episodes across 8 GPUs
+- `RYWORLD_STOP_DEBUG=1` (per-step disc_argmax + stop_prob captured in server.log.gz)
+- Eval framework: `third_party/vlnverse_emr` `VlnPeEvaluator` + `VLNPEMetric`
+- Data: `data/vlnverse/raw_data/final_splits/coarse_val_unseen.json.gz` (835 ep, raw — evaluator picks `formal` instruction on-the-fly)
+
+## Comparison vs VLNVerse paper (arXiv:2512.19021) Table 3 baselines
+
+| Method | SR | SPL | Δ vs RyWorld |
+|--------|----|----|--------------|
+| CMA (VLN-CE) | 32.15% | 29.06% | -19.0 |
+| Seq2Seq | 31.91% | 29.68% | -19.2 |
+| HNR | 36.02% | 33.67% | -15.1 |
+| RDP | 41.61% | 37.53% |  -9.5 |
+| GAMA (paper SOTA) | 42.45% | 38.89% |  -8.7 |
+| **RyWorld @ thr=0.95** | **51.14%** | **49.22%** | — |
+
+## Stop head behavior (151,740 chunk positions)
+
+| Statistic | Value |
+|-----------|-------|
+| stop_prob median | 0.752 |
+| stop_prob p90 | 0.897 |
+| pathA fire (argmax==Stop natural) | 2.51% |
+| pathB fire (threshold override) | 0.68% |
+| no-stop | 96.81% |
+
+## Bucket SR by shortest_path
+
+| shortest_path | n | SR | NE |
+|--------------|---|-----|-----|
+| [ 0,  5) m | 151 | 66.9% | 2.55 |
+| [ 5,  8) m | 360 | 59.4% | 3.07 |
+| [ 8, 12) m | 226 | 42.9% | 4.33 |
+| [12, 18) m |  96 | 14.6% | 6.58 |
+| [18, 30) m |   2 | 50.0% | 4.55 |
+
+## Files in this package
+
+- `shards/<name>/`: per-shard artifacts
+  - `meta.json` — provenance + aggregate per shard
+  - `per_episode.csv` — every episode's NE/success/SPL/etc
+  - `eval_result.log` — vlnverse_emr official format
+  - `eval.log.gz` — full eval client log
+  - `server.log.gz` — full agent server log (incl STOP_DEBUG per chunk position)
+  - `eval_cfg.py` — exact cfg used
+- `scripts/`: aggregation + comparison scripts
+  - `agg_dedup_v2.py` — dedupe-by-tid aggregator (canonical numbers)
+  - `final_verify.py` — multi-source cross-check
+  - `sweep_diff.py`, `eval_metrics2.py`, `ground_truth_analysis.py` — supplementary analyzers
+
+## Reproduction
+1. Same ckpt + same cfg + same `stop_threshold=0.95`
+2. Run `scripts/eval/run_eval_structured.sh` 8 times across GPUs with sharded data
+3. Aggregate with `agg_dedup_v2.py`

README.md

@@ -0,0 +1,211 @@
+---
+license: apache-2.0
+base_model: OpenGVLab/InternVL3_5-1B
+pipeline_tag: robotics
+library_name: pytorch
+tags:
+  - vision-language-navigation
+  - VLN
+  - VLNVerse
+  - InternVL
+  - robot-navigation
+  - sim2real
+  - embodied-AI
+model-index:
+  - name: RyWorld VLN — Stage 1 Discrete (step 15000)
+    results:
+      - task:
+          type: vision-language-navigation
+          name: VLN coarse/val_unseen
+        dataset:
+          name: VLNVerse coarse/val_unseen
+          type: Eyz/VLNVerse_data
+          split: val_unseen
+        metrics:
+          - type: success_rate
+            value: 51.14
+            name: Success Rate (%)
+          - type: spl
+            value: 49.22
+            name: SPL (%)
+          - type: oracle_success_rate
+            value: 64.79
+            name: Oracle Success Rate (%)
+          - type: navigation_error
+            value: 3.727
+            name: Navigation Error (m)
+          - type: ndtw
+            value: 0.9445
+            name: nDTW
+---
+
+# RyWorld VLN — Stage 1 Discrete (step 15000)
+
+**Vision-language navigation policy** built on InternVL3.5-1B with a separate
+StopHead and ProprioProjector. Trained on the VLNVerse coarse/fine training
+set; evaluated on the official `coarse/val_unseen` 835-episode benchmark using
+the [vlnverse_emr](https://github.com/sihaoevery/vlnverse) evaluation framework.
+
+## Headline result
+
+On the full **VLNVerse coarse/val_unseen (835 episodes)** with `stop_threshold = 0.95`:
+
+| Metric | Value |
+|--------|-------|
+| **Success Rate (SR)** | **51.14%** |
+| **SPL** | **49.22%** |
+| Oracle Success Rate (OSR) | 64.79% |
+| Navigation Error (NE) | 3.727 m |
+| nDTW | 0.9445 |
+| Mean Trajectory Length | 6.121 m |
+
+## Comparison vs VLNVerse paper baselines
+
+Reproduced inside the official `vlnverse_emr` framework on the same `coarse/val_unseen`
+split. Baseline numbers from VLNVerse paper ([arXiv:2512.19021](https://arxiv.org/abs/2512.19021), Table 3):
+
+| Method | SR ↑ | SPL ↑ | Δ vs RyWorld |
+|--------|------|-------|--------------|
+| CMA (VLN-CE) | 32.15% | 29.06% | −18.99 / −20.16 |
+| Seq2Seq | 31.91% | 29.68% | −19.23 / −19.54 |
+| HNR | 36.02% | 33.67% | −15.12 / −15.55 |
+| RDP | 41.61% | 37.53% |  −9.53 / −11.69 |
+| GAMA (paper SOTA) | 42.45% | 38.89% |  **−8.69 / −10.33** |
+| **RyWorld @ thr=0.95 (this model)** | **51.14%** | **49.22%** | — |
+
+## Architecture
+
+```
+Inputs:                              Outputs (per chunk position, chunk_size=4):
+- RGB 256x256 (Isaac live or         - Discrete head xattn: 4-way CE
+  pre-rendered training video)         (0=Stop / 1=Forward / 2=TurnLeft / 3=TurnRight)
+- Instruction text (formal variant)  - Stop head xattn: BCE-with-pos_weight
+- Proprio history N=8 keyframes        soft target stop_proximity = exp(-d/tau)
+  (body-frame deltas [dx, dy,          tau=4.33 aligned to success_radius=3 m
+   cos(dtheta), sin(dtheta)])
+- Previous action class history
+  (decision-point keyframe selector)
+
+Backbone: OpenGVLab/InternVL3_5-1B (InternViT-300M + Qwen3-0.6B, d_model=1024,
+          vision tower frozen, LoRA r=8 on language, mlp1 trainable)
+Connector: ProprioProjector (continuous proprio -> 1024 embedding)
+```
+
+Detailed architecture & training in `docs/RYWORLD_ARCHITECTURE.md` of the source repo.
+
+## Per-segment performance
+
+SR broken down by reference path length (`shortest_path_length`):
+
+| Path length (m) | n | SR | NE (m) |
+|-----------------|---|----|--------|
+| [ 0,  5) | 151 | 66.9% | 2.55 |
+| [ 5,  8) | 360 | 59.4% | 3.07 |
+| [ 8, 12) | 226 | 42.9% | 4.33 |
+| [12, 18) |  96 | 14.6% | 6.58 |
+| [18, 30) |   2 | 50.0% | 4.55 |
+
+The drop on long paths (>12 m) is the dominant remaining gap; addressing it
+likely requires either training-time long-horizon planning supervision or
+larger `forward_distance` per high-level action.
+
+## Stop head behavior (151,740 chunk-positions)
+
+| Statistic | Value |
+|-----------|-------|
+| stop_prob median | 0.752 |
+| stop_prob p90 | 0.897 |
+| pathA fire (argmax==Stop, natural) | 2.51% |
+| pathB fire (threshold override) | 0.68% |
+| no-stop | 96.81% |
+
+`stop_threshold=0.95` was selected via a 4-point sweep (0.88/0.92/0.95/0.97) on
+a 30-episode smoke subset before the full run. Higher thresholds (0.97+) cause
+overshoot regressions on the long-path segment; 0.95 is the empirical sweet spot.
+
+## How to use
+
+### 1. Load the checkpoint
+
+```python
+import torch
+from omegaconf import OmegaConf
+import sys
+sys.path.insert(0, "/path/to/ry-dynamics-vln-ryworld")
+from ryworld.training.train_ryworld_vlm import build_model_from_yaml
+from ryworld.training.ryworld_train_utils_vlm import load_vlm_checkpoint
+
+cfg = OmegaConf.merge(
+    OmegaConf.load("stage1_discrete.yaml"),
+    OmegaConf.load("a100_4gpu_discrete.yaml"),
+)
+model = build_model_from_yaml(cfg, device=torch.device("cuda"))
+load_vlm_checkpoint(model, None, "ckpt_step0015000_final.pt", strict_model=False)
+model.train(False)  # inference mode
+```
+
+### 2. Evaluate on VLNVerse
+
+```bash
+cd /path/to/ry-dynamics-vln-ryworld
+conda activate vlnverse  # Isaac Sim 4.5 + torch 2.7.1 + cu126
+export OMNI_KIT_ACCEPT_EULA=YES
+
+bash scripts/eval/run_eval_structured.sh \
+  --ckpt ckpt_step0015000_final.pt \
+  --tag eval_replicate \
+  --stop-thr 0.95
+```
+
+See `scripts/eval/run_eval_structured.sh` for the eval pipeline (records meta.json + per_episode.csv + appends to docs/eval_ledger.jsonl).
+
+## Training data
+
+- **VLNVerse coarse + fine train** (~12,000 trajectories, 33 indoor scenes)
+- Pre-rendered RGB videos at 256x256 (10 fps)
+- Discrete action labels (0=Stop / 1=Forward / 2=TurnLeft / 3=TurnRight)
+- Formal-variant instruction text
+
+Trained on 4x A100 80 GB with `chunk_size=4` multi-step CE supervision +
+StopHead BCE (pos_weight=5.0, tau=4.33).
+
+## Files in this repo
+
+| File | Description |
+|------|-------------|
+| `ckpt_step0015000_final.pt` | Main checkpoint (2.81 GB) |
+| `stage1_discrete.yaml` | Base training config |
+| `a100_4gpu_discrete.yaml` | Production overlay (4x A100) |
+| `h1_ryworld_cfg_vlnverse_coarse_val_unseen.py` | Eval config (vlnverse_emr) |
+| `eval_results/` | Full eval artifacts: per-shard meta.json, per_episode.csv, server.log.gz with STOP_DEBUG |
+| `EVAL_SUMMARY.md` | One-page summary of headline metrics |
+
+## Citation
+
+```bibtex
+@misc{ryworld2026,
+  title  = {RyWorld: Vision-Language Navigation with a Unified Multimodal World Model},
+  author = {RyWorld Team, Ruyi Dynamics},
+  year   = {2026},
+  url    = {https://huggingface.co/ruyidynamics/ryworld-vln-stage1-discrete-15k}
+}
+```
+
+If you use this model on the VLNVerse benchmark, please also cite the underlying
+benchmark paper:
+
+```bibtex
+@article{vlnverse2025,
+  title   = {VLNVerse: A Large-Scale Extensible Benchmark for Vision-Language Navigation},
+  author  = {Sihao Yu and Yuxuan Zhang and others},
+  journal = {arXiv preprint arXiv:2512.19021},
+  year    = {2025}
+}
+```
+
+## License
+
+Apache-2.0 (model weights & code).
+
+Note: VLNVerse data and Isaac Sim assets retain their own licenses (see the
+[VLNVerse repo](https://github.com/sihaoevery/vlnverse) for details).

a100_4gpu_discrete.yaml

@@ -0,0 +1,256 @@
+# A100 4×GPU Discrete training recipe — VLNVerse R2R-CE.
+#
+# 这是单文件 overlay：所有 A100 4GPU 的 batch / LR / schedule + 数据采样 +
+# inflection-weight + weight_decay 调优都在这里。原 ``stride1_iw32.yaml`` 和
+# ``_run_nocache_mlp1_train.yaml`` 已合并入此文件，避免覆盖顺序陷阱
+# （OmegaConf last-wins，多个 overlay 间冲突 key 容易静默失效）。
+#
+# **2026-05-08 默认切换为 no-cache + mlp1 trainable**：让 1M 参数的 mlp1
+# connector 学 IsaacSim 渲染 → InternViT WebData 的 sim domain gap。原
+# cache + frozen mlp1 baseline 通过 cli override 可恢复（见 §Ablation 节）。
+#
+# Recipe rationale (4 层修复，按贡献排序):
+#
+# 1. 视觉泛化 (sample_stride=1 + inflection_weight_coef=3.2)
+#    旧默认 sample_stride=3 (decision-point-aware) 只让模型见到 ~33% steady-state
+#    Forward 视觉上下文。replay-eval 在 stride=1 (closed-loop 实际分布) 上 Forward
+#    acc=35%，stride=3 (训练同分布) 70% —— 35pp gap 是结构性视觉泛化失败。
+#    fix: stride=1 让视觉覆盖每帧；iw=3.2 在 loss 里给"行动切换帧" 3.2× 权重。
+#    stop_oversample_factor=1 因为 iw=3.2 × class_weight[Stop]=3.0 已给 Stop ~9.6×
+#    有效梯度（叠 class_weights 后层级：Stop 9.6 > Turn 3.2 > F-decision 1.28 >
+#    F-steady 0.4）。vlnverse_emr 5 个 baseline 全部用 iw=3.2。
+#
+#    class_weights 重标定 [2.0,0.6,1.2,1.2] → [3.0,0.4,1.0,1.0]:
+#    base 的 [2.0,0.6,1.2,1.2] 是为 stride=3+oversample=4 自洽标定的；本 overlay 切到
+#    stride=1+oversample=1+iw=3.2 后，旧 weights 会让 Stop 梯度 share 跌到 7.6%（远低于
+#    原 21% 设计目标）。实测 per-frame 分布 1.99%/66.83%/16.54%/14.65% + 上述 iw 加权后，
+#    新 weights [3.0,0.4,1.0,1.0] 给出 Stop 14% / F 33% / TL 27.5% / TR 25.5%（温和折中，
+#    不极端到 [4.14,0.23,1.0,1.0] 完全恢复旧 21/21/30/30 target）。
+#
+# 2. 训练量 (max_steps=15000，约 9.86 epoch)
+#    实测 stride=1 数据 = 7928 trajectories × 49.1 frames = 389,499 samples/epoch。
+#    effective batch = 4 GPU × 64 = 256 → 1521.5 steps/epoch (实测精确值)。
+#    15000 / 1521.5 ≈ 9.86 epoch — 1B + LoRA fine-tune 经验甜区上沿。
+#    旧 max_steps=3000 在 stride=1 下 = 1.97 epoch（严重欠训）。
+#    vlnverse_emr 跑 80 epoch 但模型是小 CMA，不直接对照。
+#
+# 3. weight_decay (0.01 → 1e-4)
+#    base AdamW 默认 0.01 对 1B + LoRA 偏强，可能压扁 LoRA 与 scratch heads。
+#    vlnverse_emr 用 1e-5（小 CMA 可承受），我们走中点 1e-4。
+#
+# 4. 长程 spatial memory (keyframe selector_type=decision_point + max_history=8)
+#    之前 FIFO 4 是激进轻量化，三个理由经审视后均不成立或失效：(a) 闭环 inference
+#    没法 decision_point 是错误的——NavigationMemory.update 是增量计算只查
+#    yaw 与最近 KF 差，不需未来；vlm_runtime 已正确使用同 selector，无 train-infer
+#    mismatch；(b) ViT 提速 2× 在 vision_cache 启用后失效（ViT compute = 0）；
+#    (c) "Stop/Turn 短瞬态" 不解释 Forward steady-state 视觉混淆（0.4s FIFO 看
+#    不出"我已 Forward 5 步还是 12 步"），这正是 vlnverse_emr 报告的 R2R-CE 主因。
+#    8 = 1 anchor (frame 0) + 3 KF (yaw>15° 决策点) + 4 recent，覆盖长/中/短程
+#    spatial memory。文献甜点 6-8（DUET/HAMT/NavGPT/vlnverse_emr 5 baselines）。
+#    Cost: 序列长度 1.78×（attention O(N²) 增加），per-step wall 估 +30-50%；
+#    cache 让 ViT compute 持平。
+#
+#    **2026-05 实测 step time（A100 4-GPU batch=64 per-rank, N_h=8 decision_point）**:
+#      cache mode + frozen mlp1   ≈  9-10 sec/step  → ~38-42 h 总训 (15k steps)
+#      no-cache + trainable mlp1  ≈ 12-13 sec/step  → ~50-54 h 总训
+#    （旧上古 ~4 sec/step 估算来自 FIFO N_h=4 时代，已不适用 N_h=8）
+#
+# 5. 性能优化已落地（详见 docs/RYWORLD_ARCHITECTURE.md §4.12）
+#    a. forward_with_cached_vision_features_batched (单次 LM batched forward 替代 per-sample loop)
+#    b. forward_batched (no-cache 路径同样 batched)
+#    c. logits_to_keep=1 跳 lm_head 全长 GEMM (~2 sec/step + 45 GiB transient)
+#    d. batched proprio inject (索引一次替代 B*N_h 次元素赋值)
+#    e. 显式 attn_implementation="flash_attention_2"
+#    f. worker-side PIL preprocess (parallel via num_workers=16)
+#    g. drop input_embeds.clone() (代码清爽，性能微正)
+#
+# Launch (StopHead 默认 ON；**vision_cache 默认 OFF + mlp1 trainable**):
+#
+# 默认 (no-cache + mlp1 trainable, ~13.5 sec/step DDP, ~50-56 h 总训):
+#   推荐用 ./run_train.sh — 设 RYWORLD_RUN_TS 让 train.log + ckpt + run_shell.log
+#   全落到 checkpoints/{exp}_{TS}/ 同目录。Manual 等价：
+#     PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
+#     torchrun --nproc_per_node=4 -m ryworld.training.train_ryworld_vlm \
+#       --config ryworld/configs/stage1_discrete.yaml \
+#       --config-override ryworld/configs/overlays/a100_4gpu_discrete.yaml
+#
+# Ablation A (cache mode + frozen mlp1, ~9-10 sec/step DDP, ~38-42 h 总训):
+#   train_ryworld_vlm argparse 不收 cli key=value 位置参数。Ablation 用
+#   一个 2 行 overlay 文件 _ablation_cache.yaml（model.trunk.freeze_mlp1=true
+#   + data.vision_cache_root=/robot/tw/vlnverse_vision_cache_internvl1b），
+#   然后双 overlay 叠加：
+#     # 一次性预算 cache (cache-root 由 precompute 脚本参数指定，无需 yaml)：
+#     python scripts/data/precompute_vision_features.py \
+#       --config ryworld/configs/stage1_discrete.yaml \
+#       --config-override ryworld/configs/overlays/a100_4gpu_discrete.yaml \
+#       --data-root /robot/tw/vlnverse_datasets \
+#       --cache-root /robot/tw/vlnverse_vision_cache_internvl1b
+#     # 训练：
+#     torchrun ... \
+#       --config-override ryworld/configs/overlays/a100_4gpu_discrete.yaml \
+#       --config-override ryworld/configs/overlays/_ablation_cache.yaml
+#
+# 评判（用 scripts/eval/replay_offline_eval.py）:
+#   * stride=1 Forward acc 旧 35% → 目标 ≥ 60%
+#   * stride=3 Forward acc ≥ 65%
+#   * 训练曲线 classes 间 loss 差距缩小，5k-8k step 后仍下降
+
+# ---------------------------------------------------------------------------
+# Model — VEPFS 本地权重路径 + class_weights 重标定 + 默认启用 StopHead
+#         (StopHead 字段 2026-05-07 由原 overlays/stop_head.yaml 合并入此 recipe；
+#          VLNVerse goal_position 数据 100% 命中、推理路径与训练对齐已落地，
+#          StopHead 现为本 recipe 默认配置；单独 overlay 已删除以避免覆盖顺序陷阱)
+#
+#         **freeze_mlp1=false (默认 2026-05-08)**：让 1M 参数的 mlp1 connector
+#         学 IsaacSim 渲染 → InternViT WebData 的 sim domain gap。VLNVerse 的图像
+#         全部由 IsaacSim 渲染 (贴图风格 / 光照频谱 / 纹理统计)，跟 InternViT 在
+#         WebData 上 pretrain 的真实分布有不可忽视的 gap；mlp1 是唯一能在 vision
+#         side 吸收 gap 的 trainable 1M 参数 (vision tower 仍冻结)。Cost: 训练图
+#         多走 ViT (~2 sec/step) + mlp1 forward+bwd (~0.1 sec)，~+25-30% step time
+#         vs cache mode。Cache 模式 (frozen mlp1) 通过 cli override 可恢复，见
+#         上方 Ablation A 节。
+#
+#         **vision_cache_root=null (默认 2026-05-08)**：禁用 cache 配套
+#         freeze_mlp1=false。cache + freeze_mlp1=False 是 silent 死锁 (cache
+#         绕过 mlp1 forward → mlp1 trainable 但永远收不到梯度)，
+#         train_ryworld_vlm.py 启动时 guard 显式 raise 阻止该组合。
+# ---------------------------------------------------------------------------
+model:
+  trunk:
+    model_id_or_path: /imotion-vepfs/imotion_zxnet/workspace_wei/hf_models/InternVL3_5-1B
+    local_files_only: true
+    freeze_mlp1: false           # default 2026-05-08: mlp1 trainable 学 sim domain gap
+  joint:
+    # Separate binary StopHead — decouples "am I at the goal?" from
+    # navigation argmax. Soft labels via stop_proximity = exp(-d/τ); inference
+    # gate P(stop) > stop_threshold (default 0.5) overrides argmax with class 0.
+    enable_stop_head: true
+    stop_head_type: xattn
+    stop_xattn_n_layers: 2
+    stop_xattn_dropout: 0.1
+    # Action history conditioning — fuse past N keyframes' discrete action
+    # class into the per-keyframe embedding via Embedding(4, d_model). Targets
+    # the Forward steady-state visual confusion (visual context alone cannot
+    # distinguish "forwarding for 3 steps" from "forwarding for 10" in a
+    # corridor; explicit class history makes it trivial).
+    #
+    # Re-enable (2026-05-08): the 2026-05-07 triage crash on this overlay
+    # ("DDP reducer reports static_graph violation on iter 2") was caused
+    # by all-pad early-episode samples (history_mask all-False) — the
+    # per-sample slice of action_emb's batched output collapses to
+    # ``[0, d_model]``, so on the very first iter autograd can't see
+    # action_emb participating, ``static_graph=True`` locks it as
+    # "unused", and any later iter with a non-empty mask flips the
+    # participation pattern → reducer error. The longer history window
+    # under decision_point (max_history=8 vs FIFO=4) makes early-episode
+    # all-pad batches more likely. Fix lives in ``compute_losses`` of
+    # ``ryworld/models/world/ryworld_vlm_joint.py``: action_emb +
+    # action_emb_scale are now unconditionally folded into the
+    # zero-scaled dummy term whenever the module is allocated, which
+    # guarantees consistent grad participation across iters. Verified by
+    # tests/test_action_history.py + an all-pad batch smoke test.
+    use_action_history: true
+    action_emb_init_scale: 0.1
+  loss:
+    # 2026-05-10 重新标定（旧 [3.0, 0.4, 1.0, 1.0] + iw=3.2 实测 cls1 acc plateau
+    # 在 0.43-0.49，pred_frac 0.34 vs label 0.67 gap 0.32，模型完美收敛到 loss
+    # 设计的 Forward share 33%。Bayesian 反推：要 replay-eval Forward acc ≥ 60%，
+    # designed share 需 ≥ 55% → cw[1] 0.4 → 0.7（+75%）。同时 iw 3.2 → 2.5
+    # （-22%），减少 inflection 帧（多为 Turn）的 boost，间接释放 Forward share。
+    # 新 designed share: Stop 12% / Forward 53% / TurnL 20% / TurnR 18%。
+    # 历史: 旧 base [2.0,0.6,1.2,1.2] 是为 stride=3+oversample=4 标定的。
+    discrete_class_weights: [3.0, 0.7, 1.0, 1.0]
+    # StopHead BCE 与 VLNVerse success_radius=3.0m 对齐：
+    #   - stop_proximity = exp(-d/τ)。要让 d=success_radius 时软目标 ≈ 推理阈值
+    #     stop_threshold=0.5（即 d=3m 处训练-推理决策点对齐），解
+    #     τ = -3.0/ln(0.5) ≈ 4.33。
+    #     旧值 τ=2.0 让 d=3m 时目标仅 0.22 → 推理 0.5 阈值实际触发距离仅 1.39m，
+    #     模型可能穿过 success 半径仍不停，浪费 navigation steps。
+    #   - lambda_stop=0.5（保守起步）：BCE-with-pos_weight=5 数值上可能远大于
+    #     4-class focal CE，1.0 同权易让 stop loss 主导梯度让 navigation 头退化。
+    #     依据前 500 step 两 loss 数值与 stop_recall/precision 再回调。
+    #   - focal γ=2 强压大量 negative；pos_weight=5 补 Stop 帧 ~2% 类频。
+    lambda_stop: 0.5
+    stop_focal_gamma: 2.0
+    stop_proximity_tau: 4.33
+    stop_pos_weight: 5.0
+
+# ---------------------------------------------------------------------------
+# Data — A100-80GB 大 batch + 视觉泛化 fix (stride=1 + iw=3.2 + oversample=1) +
+#        decision_point 8 历史帧（1 anchor + 3 KF + 4 recent，长程+决策点+短程）+
+#        no-cache (ViT + mlp1 进训练图，mlp1 学 sim domain gap)
+#
+# Keyframe selector 选型（2026-05 决策）：
+#   * 之前 FIFO 4 的三个理由经审视后均不成立或失效：
+#     (a) 闭环 inference 没法 decision_point —— 错。NavigationMemory.update 是
+#         增量计算，只比较与最近 KF 的 yaw 差，不需要未来序列；vlm_runtime
+#         已正确使用同一个 selector，不存在 train-infer mismatch。
+#     (b) ViT 提速 2× —— 切到 no-cache 后 ViT compute 又回到训练图 ~2 sec/step，
+#         FIFO 4 vs decision_point 8 的 attention seq 长度 1.78× 差距重新有意义；
+#         但 use_action_history=true 已补偿 Forward steady-state 视觉混淆。
+#     (c) Stop/Turn 1-2 帧短瞬态够用 —— 不解释 Forward steady-state 失明
+#         （0.4s FIFO 看不出"我已 Forward 5 步还是 12 步"），这正是 vlnverse_emr
+#         报告的 R2R-CE Forward 视觉混淆主因。
+#   * 文献甜点 6-8 history（DUET / HAMT / NavGPT / vlnverse_emr 5 baselines），
+#     FIFO 4 是激进轻量化偏离。
+#   * 选 8 = 1 anchor (frame 0, episode 起点场景锚) + 3 KF (yaw 变化决策点) +
+#     4 recent (短期连续上下文)，三段 buffer 覆盖长/中/短程 spatial memory。
+#   * 序列长度 ~1320 → ~2350 (1.78×)，attention O(N²) 升 3×；no-cache 模式
+#     ViT compute 回训练图 ~2 sec/step。实测 ~12-13 sec/step DDP @ 4xA100。
+# ---------------------------------------------------------------------------
+data:
+  root: /robot/tw/vlnverse_datasets   # 7928 traj VLNVerse R2R-CE，npy 格式
+  vision_cache_root: null      # default 2026-05-08: ViT+mlp1 进训练图；
+                               # cache 模式恢复见 Ablation A 节
+  batch_size: 64               # 64 + GC=true 是经测试可装的配置；
+                               # global_batch = 4 × 64 = 256
+                               # 实测过 batch=80：no speedup (samples/sec 持平 26.5)，
+                               # GPU 已 94% compute-saturated，加 batch 只是按比例
+                               # 摊 wall-clock + 险些 OOM（72.8/80 GB）。
+                               # 加速只能从结构性维度找（vision token 降密 / N_h 砍半）。
+  num_workers: 16              # 8 imgs × 64 batch / rank = 512 frames load/step；
+                               # cache load + ProprioProjector 预算够用
+  sample_stride: 1
+  stop_oversample_factor: 1
+  inflection_weight_coef: 2.5    # 2026-05-10: 3.2 → 2.5。iw 主要 boost Turn 起止帧
+                                 # （inflection 集中在 Turn 而非 Forward steady state），
+                                 # 间接持续压 Forward share。降到 2.5 配合 cw[1]=0.7
+                                 # 共同把 designed Forward share 从 33% 抬到 53%。
+                                 # vlnverse_emr 用 3.2 但他们是小 CMA model + 不同
+                                 # class_weights 平衡，不直接对照。
+  keyframe:
+    # 8 = 1 anchor + 3 KF region + 4 recent reserved。
+    selector_type: decision_point
+    max_history: 8
+    yaw_thresh_deg: 15.0
+    min_frame_gap: 1             # VLNVerse 10fps，每帧都可成 KF；阈值由 yaw 控制
+    protect_first_frame: true    # frame 0 始终保留作为 episode anchor
+    yaw_lowpass_alpha: 0.0       # discrete env yaw 信号干净，不需低通
+    reserve_recent_frames: 4     # 最新 4 帧不 evict（短期连续上下文）
+
+# ---------------------------------------------------------------------------
+# Optim — sqrt(4)=2.0 LR scaling for batch 64→256，wd 放松到 1e-4
+# ---------------------------------------------------------------------------
+optim:
+  lr_scratch: 6.0e-4           # 3.0e-4 × 2.0
+  lr_lora: 1.0e-4              # 5.0e-5 × 2.0
+  lr_connector: 1.0e-4         # 5.0e-5 × 2.0
+  weight_decay: 1.0e-4         # base 0.01 → 1e-4
+
+# ---------------------------------------------------------------------------
+# Train — ~9.86 epochs at stride=1 (实测 1521.5 steps/epoch × 9.86 ≈ 15000)
+# ---------------------------------------------------------------------------
+train:
+  max_steps: 15000
+  lr_warmup_steps: 1500        # 10% — mlp1 newly trainable + use_action_history 首次同时生效，
+                               # 加上 batch 256 大；5% (750 step) 太激进，10% 给 mlp1 + action_emb
+                               # 充分 ramp 避免前 500 step NaN/loss spike。代价：cosine 衰减 step
+                               # 从 14250 → 13500，主体训练阶段缩 5%，可忽略。
+  lr_cosine_decay_steps: 13500 # max_steps - warmup
+  save_every: 1500             # ~每 epoch 1 次 (steps 1500, 3000, ..., 15000 = 10 ckpt 总数)
+  log_every: 10
+  gradient_checkpointing: true   # 必需，batch=64 + LM 28 层不开 GC OOM
+
+wandb:
+  run_name: stage1_discrete_4xA100_cw07_iw25
+  tags: [stage1, discrete, vlnverse, internvl, 1b, 4xA100, stride1, iw25, cw07, long_train, stophead, no_cache, mlp1_trainable, action_history, chunk4]

ckpt_step0015000_final.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70bec5b621db4c996522835368aba13e1e6192118251b9e46c956cdc86e09124
+size 2812658821

eval_results/SUMMARY.md

@@ -0,0 +1,70 @@
+# RyWorld VLNVerse Eval — Full coarse/val_unseen 835 ep
+
+## Final metrics (canonical: dedupe-by-trajectory_id across all 835 ep)
+
+| Metric | Value |
+|--------|-------|
+| **SR (success rate)** | **51.14%** |
+| **SPL** | **49.22%** |
+| **OSR (oracle success rate)** | **64.79%** |
+| **NE (navigation error, lower better)** | **3.727 m** |
+| **TL (trajectory length)** | **6.121 m** |
+| **nDTW (path alignment)** | **0.9445** |
+
+## Run config
+- Checkpoint: `ckpt_step0015000_final.pt` (stage1 discrete, InternVL3.5-1B + StopHead + ProprioProjector)
+- `stop_threshold` = **0.95** (selected after 4-point sweep 0.88/0.92/0.95/0.97 on 30-ep smoke)
+- Sharded 8-way: 7×104 + 1×107 episodes across 8 GPUs
+- `RYWORLD_STOP_DEBUG=1` (per-step disc_argmax + stop_prob captured in server.log.gz)
+- Eval framework: `third_party/vlnverse_emr` `VlnPeEvaluator` + `VLNPEMetric`
+- Data: `data/vlnverse/raw_data/final_splits/coarse_val_unseen.json.gz` (835 ep, raw — evaluator picks `formal` instruction on-the-fly)
+
+## Comparison vs VLNVerse paper (arXiv:2512.19021) Table 3 baselines
+
+| Method | SR | SPL | Δ vs RyWorld |
+|--------|----|----|--------------|
+| CMA (VLN-CE) | 32.15% | 29.06% | -19.0 |
+| Seq2Seq | 31.91% | 29.68% | -19.2 |
+| HNR | 36.02% | 33.67% | -15.1 |
+| RDP | 41.61% | 37.53% |  -9.5 |
+| GAMA (paper SOTA) | 42.45% | 38.89% |  -8.7 |
+| **RyWorld @ thr=0.95** | **51.14%** | **49.22%** | — |
+
+## Stop head behavior (151,740 chunk positions)
+
+| Statistic | Value |
+|-----------|-------|
+| stop_prob median | 0.752 |
+| stop_prob p90 | 0.897 |
+| pathA fire (argmax==Stop natural) | 2.51% |
+| pathB fire (threshold override) | 0.68% |
+| no-stop | 96.81% |
+
+## Bucket SR by shortest_path
+
+| shortest_path | n | SR | NE |
+|--------------|---|-----|-----|
+| [ 0,  5) m | 151 | 66.9% | 2.55 |
+| [ 5,  8) m | 360 | 59.4% | 3.07 |
+| [ 8, 12) m | 226 | 42.9% | 4.33 |
+| [12, 18) m |  96 | 14.6% | 6.58 |
+| [18, 30) m |   2 | 50.0% | 4.55 |
+
+## Files in this package
+
+- `shards/<name>/`: per-shard artifacts
+  - `meta.json` — provenance + aggregate per shard
+  - `per_episode.csv` — every episode's NE/success/SPL/etc
+  - `eval_result.log` — vlnverse_emr official format
+  - `eval.log.gz` — full eval client log
+  - `server.log.gz` — full agent server log (incl STOP_DEBUG per chunk position)
+  - `eval_cfg.py` — exact cfg used
+- `scripts/`: aggregation + comparison scripts
+  - `agg_dedup_v2.py` — dedupe-by-tid aggregator (canonical numbers)
+  - `final_verify.py` — multi-source cross-check
+  - `sweep_diff.py`, `eval_metrics2.py`, `ground_truth_analysis.py` — supplementary analyzers
+
+## Reproduction
+1. Same ckpt + same cfg + same `stop_threshold=0.95`
+2. Run `scripts/eval/run_eval_structured.sh` 8 times across GPUs with sharded data
+3. Aggregate with `agg_dedup_v2.py`

eval_results/scripts/agg_dedup_v2.py

@@ -0,0 +1,45 @@
+"""Dedup-by-tid aggregate. Handles both live eval*.log and finished eval*.log.gz."""
+import re
+import gzip
+import glob
+import os
+
+EVALS = "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/evals"
+pat = re.compile(
+    r'"shortest_path_length":\s*([\d.]+).*?"NE":\s*([\d.]+).*?"success":\s*([\d.]+).*?'
+    r'"osr":\s*([\d.]+).*?"TL":\s*([\d.]+).*?"spl":\s*([\d.]+).*?"ndtw":\s*([\d.]+).*?'
+    r'"trajectory_id":\s*"([^"]+)"',
+    re.S,
+)
+
+all_eps = {}
+log_paths = glob.glob(f"{EVALS}/*full835*/eval*.log") + glob.glob(f"{EVALS}/*full835*/eval*.log.gz")
+for p in sorted(log_paths):
+    if p.endswith(".gz"):
+        content = gzip.open(p, "rt").read()
+    else:
+        content = open(p).read()
+    for m in pat.finditer(content):
+        spg, ne, succ, osr, tl, spl, ndtw, tid = m.groups()
+        all_eps[tid] = {
+            "spg": float(spg), "NE": float(ne),
+            "success": int(float(succ) == 1.0),
+            "osr": int(float(osr) == 1.0),
+            "SPL": float(spl), "nDTW": float(ndtw),
+        }
+
+n = len(all_eps)
+if n == 0:
+    print("no eps found")
+    raise SystemExit
+eps = list(all_eps.values())
+def avg(field):
+    return sum(e[field] for e in eps) / n
+
+print(f"=== TOTAL {n}/835 ({100*n/835:.1f}%) ===")
+print(f"  SR={100*avg('success'):.1f}%  OSR={100*avg('osr'):.1f}%  NE={avg('NE'):.2f}  SPL={100*avg('SPL'):.1f}%  nDTW={avg('nDTW'):.3f}")
+for lo, hi in [(0, 5), (5, 8), (8, 12), (12, 18)]:
+    bk = [e for e in eps if lo <= e["spg"] < hi]
+    if bk:
+        b_sr = sum(e["success"] for e in bk) / len(bk)
+        print(f"  short [{lo:>2},{hi:>2})m: n={len(bk):>3}  SR={100*b_sr:>4.1f}%")

eval_results/scripts/eval_metrics2.py

@@ -0,0 +1,82 @@
+"""Per-episode metric extraction (NE / TL / SPL / OSR / nDTW / shortest_path)."""
+import re
+import sys
+
+log_path = sys.argv[1]
+log = open(log_path).read()
+
+# Per-episode JSON-like block — capture ALL metrics
+pat = re.compile(
+    r'"shortest_path_length":\s*([\d.]+).*?'
+    r'"NE":\s*([\d.]+).*?'
+    r'"success":\s*([\d.]+).*?'
+    r'"osr":\s*([\d.]+).*?'
+    r'"TL":\s*([\d.]+).*?'
+    r'"spl":\s*([\d.]+).*?'
+    r'"ndtw":\s*([\d.]+).*?'
+    r'"steps":\s*(\d+).*?'
+    r'"trajectory_id":\s*"([^"]+)".*?'
+    r'"fail_reason":\s*"([^"]+)"',
+    re.S,
+)
+
+finishes = re.compile(
+    r"\[(\d+)/30\]\[step_index:(\d+)\] finish: \[trajectory_id:([^\]]+)\]"
+    r"\[duration:([\d.]+) s\]\[step_count:(\d+)\]\[fps:([\d.]+)\]\[result:([^\]]+)\]"
+)
+
+per = {}
+for m in pat.finditer(log):
+    spl_geo, ne, succ, osr, tl, spl, ndtw, steps, tid, reason = m.groups()
+    per[tid] = dict(
+        spl_geo=float(spl_geo), NE=float(ne), success=float(succ),
+        osr=float(osr), TL=float(tl), SPL=float(spl), nDTW=float(ndtw),
+        steps=int(steps), reason=reason,
+    )
+
+H = "#  | trajectory_id            | result          |   NE | shortP |   TL |  SPL |  OSR | nDTW | steps |  sec | fps "
+print(H)
+print("-" * len(H))
+agg = {k: 0.0 for k in ["NE", "TL", "SPL", "OSR", "nDTW", "spl_geo"]}
+succ_n = 0
+n = 0
+fail_tally = {}
+for m in finishes.finditer(log):
+    idx, step_idx, tid_full, dur, stepc, fps, result = m.groups()
+    parts = tid_full.split("_")
+    if len(parts) >= 8 and "_".join(parts[: len(parts) // 2]) == "_".join(parts[len(parts) // 2 :]):
+        tid_key = "_".join(parts[: len(parts) // 2])
+    else:
+        tid_key = tid_full
+    e = per.get(tid_key) or {}
+    NE = e.get("NE", float("nan"))
+    TL = e.get("TL", float("nan"))
+    SPL = e.get("SPL", float("nan"))
+    OSR = e.get("osr", float("nan"))
+    nDTW = e.get("nDTW", float("nan"))
+    spg = e.get("spl_geo", float("nan"))
+    succ = e.get("success", 1.0 if result == "success" else 0.0)
+    reason = e.get("reason", result)
+    n += 1
+    succ_n += int(succ == 1.0)
+    fail_tally[reason] = fail_tally.get(reason, 0) + 1
+    for k, v in [("NE", NE), ("TL", TL), ("SPL", SPL), ("OSR", OSR), ("nDTW", nDTW), ("spl_geo", spg)]:
+        if v == v:
+            agg[k] += v
+    print(
+        f"{idx:>2} | {tid_key[:24]:<24} | {result[:15]:<15} | "
+        f"{NE:>4.2f} | {spg:>6.2f} | {TL:>4.2f} | {SPL:>4.2f} | "
+        f"{OSR:>4.2f} | {nDTW:>4.2f} | {stepc:>5} | {float(dur):>4.0f} | {float(fps):>4.1f}"
+    )
+
+print("-" * len(H))
+if n:
+    print(f"\nAggregates over {n} episodes:")
+    print(f"  SR    = {succ_n}/{n} = {100*succ_n/n:5.1f}%")
+    print(f"  NE    = {agg['NE']/n:5.2f} m  (lower better; success_radius = 3.0 m)")
+    print(f"  SPL   = {agg['SPL']/n:5.3f}    (success-weighted, lower path = better)")
+    print(f"  OSR   = {agg['OSR']/n:5.3f}    (oracle success rate — could have stopped within radius at any step)")
+    print(f"  nDTW  = {agg['nDTW']/n:5.3f}    (path alignment 0-1, higher better)")
+    print(f"  TL    = {agg['TL']/n:5.2f} m  (mean trajectory length)")
+    print(f"  shortest_path = {agg['spl_geo']/n:5.2f} m (mean geodesic)")
+    print(f"  fail_reasons: {fail_tally}")

eval_results/scripts/final_verify.py

@@ -0,0 +1,118 @@
+"""Final verification of full 835-ep eval metrics — cross-check from multiple sources."""
+import gzip
+import json
+import re
+import glob
+import os
+from collections import Counter
+
+EVALS = "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/evals"
+
+# === SOURCE 1: per-shard meta.json (vlnverse_emr's own aggregate per shard) ===
+print("=" * 90)
+print("SOURCE 1: per-shard meta.json (vlnverse_emr 自带 aggregate)")
+print("=" * 90)
+meta_files = sorted(glob.glob(f"{EVALS}/*full835*/meta.json"))
+shard_aggs = []
+for mf in meta_files:
+    m = json.load(open(mf))
+    agg = m.get("results_aggregate", {})
+    if agg.get("n_total"):
+        name = mf.split("/")[-2].split("full835_thr095_")[1] if "full835_thr095_" in mf else "?"
+        shard_aggs.append((name, agg, m.get("stop_head_stats", {})))
+        print(f"  {name:<26}  n={agg['n_total']:>3}  SR={agg['SR']*100:>5.1f}%  SPL={agg['SPL']*100:>5.1f}%  OS={agg['OS']*100:>5.1f}%  NE={agg['NE']:.2f}")
+
+# Compute weighted aggregate across shards
+total_n = sum(a[1]["n_total"] for a in shard_aggs)
+def weighted(field):
+    return sum(a[1][field] * a[1]["n_total"] for a in shard_aggs) / total_n
+print("-" * 90)
+print(f"  WEIGHTED ACROSS SHARDS  n={total_n}  "
+      f"SR={weighted('SR')*100:.2f}%  SPL={weighted('SPL')*100:.2f}%  "
+      f"OS={weighted('OS')*100:.2f}%  NE={weighted('NE'):.3f}")
+print()
+
+# === SOURCE 2: dedupe-by-trajectory_id across all eval.log files ===
+print("=" * 90)
+print("SOURCE 2: dedupe-by-trajectory_id 从所有 eval.log (含 .gz / _resume)")
+print("=" * 90)
+pat = re.compile(
+    r'"shortest_path_length":\s*([\d.]+).*?"NE":\s*([\d.]+).*?"success":\s*([\d.]+).*?'
+    r'"osr":\s*([\d.]+).*?"TL":\s*([\d.]+).*?"spl":\s*([\d.]+).*?"ndtw":\s*([\d.]+).*?'
+    r'"trajectory_id":\s*"([^"]+)"',
+    re.S,
+)
+all_eps = {}
+log_paths = sorted(glob.glob(f"{EVALS}/*full835*/eval*.log") + glob.glob(f"{EVALS}/*full835*/eval*.log.gz"))
+for p in log_paths:
+    if p.endswith(".gz"):
+        content = gzip.open(p, "rt").read()
+    else:
+        content = open(p).read()
+    for m in pat.finditer(content):
+        spg, ne, succ, osr, tl, spl, ndtw, tid = m.groups()
+        all_eps[tid] = {
+            "spg": float(spg), "NE": float(ne),
+            "success": int(float(succ) == 1.0),
+            "osr": int(float(osr) == 1.0),
+            "TL": float(tl), "SPL": float(spl), "nDTW": float(ndtw),
+        }
+
+n = len(all_eps)
+eps = list(all_eps.values())
+def avg(f):
+    return sum(e[f] for e in eps) / n
+print(f"  unique trajectory_id  n={n}/835  "
+      f"SR={100*avg('success'):.2f}%  SPL={100*avg('SPL'):.2f}%  "
+      f"OSR={100*avg('osr'):.2f}%  NE={avg('NE'):.3f}  TL={avg('TL'):.3f}  nDTW={avg('nDTW'):.4f}")
+print()
+
+# === SOURCE 3: vlnverse_emr's eval_result.log (official format) per shard ===
+print("=" * 90)
+print("SOURCE 3: vlnverse_emr eval_result.log (官方格式)")
+print("=" * 90)
+result_logs = sorted(glob.glob(f"{EVALS}/*full835*/eval_result.log"))
+for rl in result_logs:
+    name = rl.split("/")[-2].split("full835_thr095_")[1] if "full835_thr095_" in rl else "?"
+    content = open(rl).read()
+    # Parse SR/SPL/NE
+    sr_m = re.search(r"SR = ([\d.]+) / (\d+) = ([\d.]+)%", content)
+    sp_m = re.search(r"SPL = ([\d.]+) / (\d+) = ([\d.]+)%", content)
+    if sr_m and sp_m:
+        print(f"  {name:<26}  SR={sr_m.group(1):>5}/{sr_m.group(2)} = {sr_m.group(3)}%  "
+              f"SPL={sp_m.group(3)}%")
+print()
+
+# === SOURCE 4: stop_head_stats across shards ===
+print("=" * 90)
+print("SOURCE 4: stop_head_stats(模型 stop head 行为)")
+print("=" * 90)
+all_sh = []
+for name, agg, sh in shard_aggs:
+    if sh.get("n_samples"):
+        all_sh.append(sh)
+        print(f"  {name:<26}  n_samples={sh['n_samples']:>5}  "
+              f"sp_p50={sh['stop_prob_p50']:.3f}  sp_p90={sh['stop_prob_p90']:.3f}  "
+              f"pA={sh['fire_rate_pathA']*100:.2f}%  pB={sh['fire_rate_pathB']*100:.2f}%  "
+              f"no_stop={sh['no_stop_rate']*100:.2f}%")
+# Weighted aggregate
+if all_sh:
+    tot = sum(s["n_samples"] for s in all_sh)
+    def wsh(f):
+        return sum(s[f] * s["n_samples"] for s in all_sh) / tot
+    print("-" * 90)
+    print(f"  WEIGHTED (n_samples={tot:>5})  sp_p50={wsh('stop_prob_p50'):.3f}  "
+          f"sp_p90={wsh('stop_prob_p90'):.3f}  pA={wsh('fire_rate_pathA')*100:.2f}%  "
+          f"pB={wsh('fire_rate_pathB')*100:.2f}%  no_stop={wsh('no_stop_rate')*100:.2f}%")
+print()
+
+# === FINAL — bucket SR ===
+print("=" * 90)
+print("FINAL — bucket SR by shortest_path")
+print("=" * 90)
+for lo, hi in [(0, 5), (5, 8), (8, 12), (12, 18), (18, 30)]:
+    bk = [e for e in eps if lo <= e["spg"] < hi]
+    if bk:
+        sr_b = sum(e["success"] for e in bk) / len(bk)
+        ne_b = sum(e["NE"] for e in bk) / len(bk)
+        print(f"  short [{lo:>2},{hi:>2})m  n={len(bk):>4}  SR={100*sr_b:>5.1f}%  NE={ne_b:.2f}")

eval_results/scripts/ground_truth_analysis.py

@@ -0,0 +1,74 @@
+"""Ground-truth per-episode agent.step analysis from eval.log."""
+import re
+import sys
+from collections import Counter
+
+eval_log = open(sys.argv[1]).read()
+
+# Pattern for actual executed action (only count the unique one, not the dup INFO line)
+# Each agent.step produces 2 lines with same timestamp; use the one without file path
+action_re = re.compile(
+    r"^\[(\d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2},\d{3})\]\[INFO\] get \d+ actions :\[\[(-?\d+)\]\]",
+    re.M,
+)
+actions = [(m.group(1), int(m.group(2))) for m in action_re.finditer(eval_log)]
+print(f"Total agent.step calls: {len(actions)}")
+
+# Episode boundaries
+finishes = list(
+    re.finditer(
+        r"\[(\d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2},\d{3})\].*?\[(\d+)/30\]\[step_index:(\d+)\] "
+        r"finish: \[trajectory_id:(\S+?)\]\[duration:([\d.]+) s\]\[step_count:(\d+)\]\[fps:([\d.]+)\]\[result:(\w+)\]",
+        eval_log,
+    )
+)
+print(f"Episodes finished: {len(finishes)}")
+
+# For each episode (1-indexed), assign actions whose timestamp <= episode finish timestamp
+# but > previous episode finish timestamp
+boundaries = [m.group(1) for m in finishes]  # finish timestamps
+ep_info = [(m.group(2), m.group(4), m.group(6), m.group(8), m.group(5)) for m in finishes]
+# (idx, tid, step_count, result, duration)
+
+def in_episode(action_ts, ep_idx):
+    """action_ts > prev_finish_ts AND action_ts <= this_finish_ts"""
+    if ep_idx == 0:
+        return action_ts <= boundaries[0]
+    return boundaries[ep_idx - 1] < action_ts <= boundaries[ep_idx]
+
+# bucket
+ep_actions = {i: [] for i in range(len(finishes))}
+for ts, cls in actions:
+    for i in range(len(finishes)):
+        if in_episode(ts, i):
+            ep_actions[i].append(cls)
+            break
+
+print()
+print(f"{'ep':>3} {'tid':<25} {'result':<14} {'sim_step':>9} {'agent_step':>10} {'first_Stop':>11} {'F%':>5} {'TL%':>5} {'TR%':>5} {'Stop%':>5}")
+print("-" * 125)
+
+for i, (idx, tid, sim_step, result, dur) in enumerate(ep_info):
+    acts = ep_actions[i]
+    n = len(acts)
+    if n == 0:
+        print(f"{idx:>3} {tid[:25]:<25} {result[:14]:<14}  NO ACTIONS")
+        continue
+    c = Counter(acts)
+    pct = lambda k: 100 * c.get(k, 0) / n
+    # first Stop position
+    first_stop = next((j for j, a in enumerate(acts) if a == 0), None)
+    fst = f"{first_stop+1}/{n}" if first_stop is not None else "never"
+    short = "_".join(tid.split("_")[: len(tid.split("_")) // 2])
+    print(
+        f"{idx:>3} {short[:25]:<25} {result[:14]:<14} {sim_step:>9} {n:>10} {fst:>11} "
+        f"{pct(1):>5.1f} {pct(2):>5.1f} {pct(3):>5.1f} {pct(0):>5.1f}"
+    )
+
+# How many actions came AFTER the last episode finish (in-progress episode)
+last_finish_ts = boundaries[-1] if boundaries else None
+in_progress = [c for ts, c in actions if last_finish_ts is None or ts > last_finish_ts]
+print(f"\n[in-progress episode]: {len(in_progress)} agent.step calls so far")
+if in_progress:
+    c = Counter(in_progress)
+    print(f"  class dist: {dict(c)}")

eval_results/scripts/sweep_diff.py

@@ -0,0 +1,51 @@
+"""4-way per-episode comparison across stop_thr sweep."""
+import re
+
+def parse(p):
+    s = open(p).read()
+    pat = re.compile(
+        r'"NE":\s*([\d.]+).*?"success":\s*([\d.]+).*?"TL":\s*([\d.]+).*?"trajectory_id":\s*"([^"]+)"',
+        re.S,
+    )
+    return {m.group(4): (float(m.group(1)), float(m.group(2)), float(m.group(3))) for m in pat.finditer(s)}
+
+import glob
+runs = {}
+for tag, glob_pat in [
+    ("v5_088", "/tmp/*v5_stopdebug.log"),
+    ("v7_092", "/tmp/*thr092_v7.log"),
+    ("v6_095", "/tmp/*v6_stopdebug.log"),
+    ("v8_097", "/tmp/*thr097_v8.log"),
+]:
+    files = glob.glob(glob_pat)
+    if not files:
+        print(f"MISSING {tag}: {glob_pat}")
+        continue
+    runs[tag] = parse(files[0])
+
+# Use v6 tids as canonical (all runs have same 30 episodes)
+keys = list(next(iter(runs.values())).keys())
+
+print(f"{'tid':<20} | {'v5_088':>10} | {'v7_092':>10} | {'v6_095':>10} | {'v8_097':>10} | sweep pattern")
+print("-" * 110)
+patterns = {}
+for k in keys:
+    cells = []
+    succ_str = ""
+    for tag in ["v5_088", "v7_092", "v6_095", "v8_097"]:
+        if k in runs[tag]:
+            ne, s, tl = runs[tag][k]
+            cells.append(f"NE{ne:>4.1f}{'✓' if s==1 else 'x'}")
+            succ_str += "1" if s == 1 else "0"
+        else:
+            cells.append(" "*10)
+            succ_str += "?"
+    patterns.setdefault(succ_str, []).append(k)
+    print(f"{k[:20]:<20} | {cells[0]:>10} | {cells[1]:>10} | {cells[2]:>10} | {cells[3]:>10} | {succ_str}")
+
+print()
+print("===pattern frequency (v5_v7_v6_v8 success bits)===")
+labels = {"1111":"always succ","0000":"always fail","0010":"only v6","0011":"v6+v8",
+          "0110":"v7+v6","0100":"only v7","0111":"v7+v6+v8","1110":"v5+v7+v6","1111":"all"}
+for pat, eps in sorted(patterns.items(), key=lambda x: -len(x[1])):
+    print(f"  {pat}: {len(eps)} ep  ({labels.get(pat,'')})  {eps[:5]}{'...' if len(eps)>5 else ''}")

eval_results/shards/shard0_serial/eval.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fa04277ee590749b17d741b96d33616d3563d2d95d89c18b6d2f02cb2a2d8f4
+size 9206580

eval_results/shards/shard0_serial/eval_cfg.py

@@ -0,0 +1,105 @@
+"""RyWorld eval cfg — VLNVerse coarse / val_unseen, H1 robot at 256x256.
+
+Forked from
+``third_party/vlnverse_emr/scripts/eval/configs/h1_cma_clip_cfg_vlnverse_coarse.py``
+with the minimum diffs needed for our model:
+
+  - ``agent.model_name='ryworld'`` so Agent.init dispatches to RyWorldAgent.
+  - ``split_data_types=['val_unseen']`` only (CMA cfg also includes val_seen).
+  - ``task_name`` carries our run identity for ``logs/<task>/`` separation.
+
+Everything else (camera 256x256, ``torso_link/h1_pano_camera_0``,
+``h1_vln_pointcloud.usd``, ``vln_pe`` env, ``flash`` mode, max_step=500,
+formal instruction) is kept identical to the CMA-CLIP baseline so our
+numbers go on the same table.
+
+Plugin registration: this cfg's top-level ``import
+ryworld.eval.vlnverse_plugin`` triggers ``@Agent.register('ryworld')`` as a
+side effect, which is what lets vlnverse_emr's dispatcher find our agent
+without a fork. ``importlib.util.spec_from_file_location`` (used by both
+``server.py`` and ``eval.py``) executes this module top-to-bottom, so by the
+time ``Agent.init`` runs the registration is already in place.
+"""
+
+from __future__ import annotations
+
+# ---------------------------------------------------------------------------
+# 1. Make the RyWorld package importable.
+#    vlnverse_emr's launchers run from inside ``third_party/vlnverse_emr``
+#    (their ``cwd``), so ``ryworld.*`` is not on ``sys.path`` by default. We
+#    walk up from this file: cfg -> configs -> vlnverse_plugin -> eval ->
+#    ryworld -> repo root.
+# ---------------------------------------------------------------------------
+import sys
+from pathlib import Path
+
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+# ---------------------------------------------------------------------------
+# 2. Import the plugin package — side effect: ``@Agent.register('ryworld')``.
+# ---------------------------------------------------------------------------
+import ryworld.eval.vlnverse_plugin  # noqa: E402, F401
+
+# ---------------------------------------------------------------------------
+# 3. Build the eval cfg using vlnverse_emr's pydantic config classes.
+# ---------------------------------------------------------------------------
+from vlnverse.configs.agent import AgentCfg  # type: ignore[import-not-found]  # noqa: E402
+from vlnverse.configs.evaluator import (  # type: ignore[import-not-found]  # noqa: E402
+    EnvCfg,
+    EvalCfg,
+    EvalDatasetCfg,
+    SceneCfg,
+    TaskCfg,
+)
+
+eval_cfg = EvalCfg(
+    agent=AgentCfg(
+        # Server port — keep clear of CMA(8087)/N1(8028) so we can run side-by-side.
+        server_port=8091,
+        model_name="ryworld",
+        # Dummy skeleton agent ignores ckpt_path; real wrapper will load this.
+        ckpt_path="/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+        model_settings={
+            "stop_threshold_override": 0.95,
+            "train_config_path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/stage1_discrete.yaml",
+            "train_config_overrides": ["/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/overlays/a100_4gpu_discrete.yaml"],
+        },
+    ),
+    env=EnvCfg(
+        env_type="vln_pe",
+        env_settings={
+            "use_fabric": False,
+            "headless": True,
+        },
+    ),
+    task=TaskCfg(
+        task_name="20260513T032330_step0015000_stop0.95_full835_thr095_shard0_serial",
+        task_settings={
+            "env_num": 1,
+            "use_distributed": False,
+            "proc_num": 1,
+            "max_step": 500,
+            "warm_up_step": 500,
+        },
+        scene=SceneCfg(
+            scene_type="kujiale",
+            scene_data_dir="data/scene_data/vlnverse",
+        ),
+        robot_name="h1",
+        robot_flash=True,
+        robot_usd_path="data/Embodiments/vln-pe/h1/h1_vln_pointcloud.usd",
+        camera_resolution=[256, 256],
+        camera_prim_path="torso_link/h1_pano_camera_0",
+    ),
+    dataset=EvalDatasetCfg(
+        dataset_type="kujiale",
+        dataset_settings={
+            "base_data_dir": "data/vlnverse/raw_data/vlnverse/coarse_shard0",
+            "split_data_types": ["val_unseen"],
+            "filter_stairs": False,
+            "instruction_type": "formal",
+        },
+    ),
+)

eval_results/shards/shard0_serial/eval_result.log

@@ -0,0 +1,16 @@
+[split:val_unseen] total 104 data
+############[val_unseen]#############
+TL = 619.2835629124288 / 104 = 5.9546
+NE = 444.2822084470055 / 104 = 4.2719
+FR = 0 / 104 = 0.0%
+StR = 0 / 104 = 0%
+OS = 63.0 / 104 = 60.58%
+SR = 45.0 / 104 = 43.269999999999996%
+SPL = 43.14515983864238 / 104 = 41.49%
+detail:
+[reach_goal]:45
+[not_reach_goal]:58
+[success]:45
+[exceed_total_max_step]:1
+[fall]:0
+##########################

eval_results/shards/shard0_serial/meta.json

@@ -0,0 +1,161 @@
+{
+  "eval_id": "20260513T032330_step0015000_stop0.95_full835_thr095_shard0_serial",
+  "tag": "full835_thr095_shard0_serial",
+  "timestamp_recorded": "2026-05-13T16:48:26+0800",
+  "ckpt": {
+    "path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+    "step": 15000,
+    "size_bytes": 2812658821
+  },
+  "code_state": {
+    "git_hash": "a4fa65370358d6f251e158f03aced7d30f167da6",
+    "branch": "feat/batched-trunk-forward",
+    "dirty": true,
+    "dirty_files": [
+      "docs/eval_ledger.jsonl",
+      "run_train.sh",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD0_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD1_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD2_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD3_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD4_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD5_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD6_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD7_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu0_thr091_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu1_thr092_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu2_thr093_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu3_thr094_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu4_thr095_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu5_thr096_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu6_thr097_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu7_thr098_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v7_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v8_port8092.py",
+      "scripts/eval/run_eval_structured.sh",
+      "third_party/vlnverse_emr",
+      "yworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen.py"
+    ],
+    "dirty_patch_md5": "aa7101e2b529b2f5596267cd1e8073ee"
+  },
+  "eval_config": {
+    "framework": "vlnverse_emr",
+    "base_cfg": "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD0_port8091.py",
+    "derived_cfg": "checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/evals/20260513T032330_step0015000_stop0.95_full835_thr095_shard0_serial/eval_cfg.py",
+    "stop_threshold": 0.95,
+    "task_name": "20260513T032330_step0015000_stop0.95_full835_thr095_shard0_serial"
+  },
+  "data_state": {
+    "scenes_used": [
+      "kujiale_0234",
+      "kujiale_0240",
+      "kujiale_0243",
+      "kujiale_0258",
+      "kujiale_0272"
+    ],
+    "scenes_with_hdr": [
+      "kujiale_0234",
+      "kujiale_0240",
+      "kujiale_0243",
+      "kujiale_0258",
+      "kujiale_0272"
+    ],
+    "scenes_missing_hdr": []
+  },
+  "hardware": {
+    "host": "di-20260502152021-n5hs7",
+    "gpus": [
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      }
+    ],
+    "count": 8
+  },
+  "results_aggregate": {
+    "TL": 5.9546,
+    "NE": 4.2719,
+    "FR": 0.0,
+    "StR": 0.0,
+    "OS": 0.6058,
+    "SR": 0.4327,
+    "SPL": 0.41490000000000005,
+    "n_reach_goal": 45,
+    "n_not_reach_goal": 58,
+    "n_success": 45,
+    "n_fall": 0,
+    "n_total": 104
+  },
+  "results_per_scene": {
+    "kujiale_0272": {
+      "n": 18,
+      "SR": 0.5555555555555556,
+      "avg_NE": 3.599298205680091,
+      "avg_SPL": 0.5367849557682614
+    },
+    "kujiale_0258": {
+      "n": 25,
+      "SR": 0.44,
+      "avg_NE": 3.6749591203982925,
+      "avg_SPL": 0.42728988742908713
+    },
+    "kujiale_0243": {
+      "n": 25,
+      "SR": 0.32,
+      "avg_NE": 5.001239707522191,
+      "avg_SPL": 0.31368471187851943
+    },
+    "kujiale_0240": {
+      "n": 13,
+      "SR": 0.6153846153846154,
+      "avg_NE": 2.925531341617832,
+      "avg_SPL": 0.5496686220824174
+    },
+    "kujiale_0234": {
+      "n": 22,
+      "SR": 0.3181818181818182,
+      "avg_NE": 5.587479246922452,
+      "avg_SPL": 0.3096806165932766
+    }
+  },
+  "stop_head_stats": {
+    "n_samples": 16028,
+    "stop_prob_min": 0.286,
+    "stop_prob_p50": 0.773,
+    "stop_prob_p90": 0.906,
+    "stop_prob_max": 0.976,
+    "stop_prob_mean": 0.7619071624656891,
+    "fire_rate_pathA": 0.023146992762665335,
+    "fire_rate_pathB": 0.010107312203643623,
+    "no_stop_rate": 0.966745695033691
+  },
+  "n_episodes_recorded": 103
+}

eval_results/shards/shard0_serial/per_episode.csv

@@ -0,0 +1,104 @@
+episode_id,scene,NE,success,osr,SPL
+kujiale_0272_31_1,kujiale_0272,0.9565219373211652,1,1.0,1.0
+kujiale_0272_4_4,kujiale_0272,1.6831183946043111,1,1.0,1.0
+kujiale_0272_43_3,kujiale_0272,1.4148630777294384,1,1.0,1.0
+kujiale_0272_7_0,kujiale_0272,2.440151297354148,1,1.0,0.6621292038287031
+kujiale_0272_36_2,kujiale_0272,2.034643769991008,1,1.0,1.0
+kujiale_0272_59_2,kujiale_0272,0.519073960960658,1,1.0,1.0
+kujiale_0272_3_2,kujiale_0272,6.072019199294342,0,0.0,0.0
+kujiale_0272_2_4,kujiale_0272,7.073432588315052,0,0.0,0.0
+kujiale_0272_33_2,kujiale_0272,6.2934146371963315,0,0.0,0.0
+kujiale_0272_53_1,kujiale_0272,2.9482114558256893,1,1.0,1.0
+kujiale_0272_55_0,kujiale_0272,1.9852838779318491,1,1.0,1.0
+kujiale_0272_1_0,kujiale_0272,6.0525158474797776,0,0.0,0.0
+kujiale_0272_52_2,kujiale_0272,0.32210834631185087,1,1.0,1.0
+kujiale_0272_38_4,kujiale_0272,8.81373594911436,0,0.0,0.0
+kujiale_0272_9_0,kujiale_0272,3.066318503786455,0,1.0,0.0
+kujiale_0272_11_0,kujiale_0272,4.691432707538336,0,1.0,0.0
+kujiale_0272_32_1,kujiale_0272,1.4910520555997604,1,1.0,1.0
+kujiale_0272_21_3,kujiale_0272,6.929470095887106,0,0.0,0.0
+kujiale_0258_6_3,kujiale_0258,1.9466750966547144,1,1.0,1.0
+kujiale_0258_21_2,kujiale_0258,1.1352773730914674,1,1.0,1.0
+kujiale_0258_1_2,kujiale_0258,5.711282352862861,0,0.0,0.0
+kujiale_0258_15_2,kujiale_0258,2.362092978299662,1,1.0,1.0
+kujiale_0258_18_3,kujiale_0258,5.735796570024189,0,0.0,0.0
+kujiale_0258_40_0,kujiale_0258,5.749722760089638,0,0.0,0.0
+kujiale_0258_4_4,kujiale_0258,5.452340782464308,0,0.0,0.0
+kujiale_0258_15_3,kujiale_0258,3.126501567766687,0,1.0,0.0
+kujiale_0258_2_3,kujiale_0258,6.345846526124659,0,1.0,0.0
+kujiale_0258_22_4,kujiale_0258,6.0179598018711395,0,0.0,0.0
+kujiale_0258_31_0,kujiale_0258,0.9521952307963624,1,1.0,1.0
+kujiale_0258_30_1,kujiale_0258,0.4642318480780592,1,1.0,1.0
+kujiale_0258_10_3,kujiale_0258,2.0941885779508955,1,1.0,1.0
+kujiale_0258_38_1,kujiale_0258,2.590535769836942,1,1.0,1.0
+kujiale_0258_38_2,kujiale_0258,5.091851969034285,0,1.0,0.0
+kujiale_0258_13_3,kujiale_0258,1.0985134114451542,1,1.0,1.0
+kujiale_0258_17_3,kujiale_0258,3.1091810796113135,0,1.0,0.0
+kujiale_0258_19_3,kujiale_0258,8.13038548626392,0,0.0,0.0
+kujiale_0258_36_4,kujiale_0258,2.296143410597163,1,1.0,1.0
+kujiale_0258_1_1,kujiale_0258,2.6128716291657583,1,1.0,1.0
+kujiale_0258_39_1,kujiale_0258,3.762097403205405,0,1.0,0.0
+kujiale_0258_12_1,kujiale_0258,4.058406198178584,0,1.0,0.0
+kujiale_0258_27_2,kujiale_0258,3.6777519868206534,0,0.0,0.0
+kujiale_0258_23_4,kujiale_0258,5.4389163824255435,0,0.0,0.0
+kujiale_0258_3_2,kujiale_0258,2.9132118172979555,1,1.0,0.6822471857271786
+kujiale_0243_37_4,kujiale_0243,7.442275451455108,0,0.0,0.0
+kujiale_0243_1_1,kujiale_0243,8.1795585637213,0,0.0,0.0
+kujiale_0243_40_3,kujiale_0243,5.001324980954627,0,1.0,0.0
+kujiale_0243_65_1,kujiale_0243,5.766108454211463,0,0.0,0.0
+kujiale_0243_49_1,kujiale_0243,3.25895933719464,0,1.0,0.0
+kujiale_0243_24_2,kujiale_0243,4.023698559146204,0,1.0,0.0
+kujiale_0243_68_2,kujiale_0243,8.656889251646744,0,0.0,0.0
+kujiale_0243_0_0,kujiale_0243,8.740760272296772,0,1.0,0.0
+kujiale_0243_7_3,kujiale_0243,7.902092337028915,0,0.0,0.0
+kujiale_0243_39_1,kujiale_0243,7.748290374858,0,0.0,0.0
+kujiale_0243_4_2,kujiale_0243,2.2141449693605013,1,1.0,1.0
+kujiale_0243_66_2,kujiale_0243,6.802335114749344,0,0.0,0.0
+kujiale_0243_56_4,kujiale_0243,0.8990432413840336,1,1.0,1.0
+kujiale_0243_23_2,kujiale_0243,0.34546759428298784,1,1.0,0.9640774053291279
+kujiale_0243_84_0,kujiale_0243,0.9822169946125936,1,1.0,1.0
+kujiale_0243_12_3,kujiale_0243,7.138812462847378,0,0.0,0.0
+kujiale_0243_57_1,kujiale_0243,0.9444424027727343,1,1.0,1.0
+kujiale_0243_61_1,kujiale_0243,4.851292863610986,0,1.0,0.0
+kujiale_0243_47_3,kujiale_0243,2.5482560234025717,1,1.0,1.0
+kujiale_0243_55_1,kujiale_0243,7.475054042165242,0,0.0,0.0
+kujiale_0243_8_4,kujiale_0243,8.580482283833884,0,0.0,0.0
+kujiale_0243_70_1,kujiale_0243,2.688535479755903,1,1.0,1.0
+kujiale_0243_79_1,kujiale_0243,5.314413186375301,0,0.0,0.0
+kujiale_0243_46_4,kujiale_0243,5.737265920516666,0,0.0,0.0
+kujiale_0243_85_3,kujiale_0243,1.7892725258708428,1,1.0,0.8780403916338572
+kujiale_0240_21_3,kujiale_0240,4.596709163692362,0,1.0,0.0
+kujiale_0240_49_3,kujiale_0240,1.8402284423170312,1,1.0,1.0
+kujiale_0240_56_0,kujiale_0240,4.058855758699873,0,1.0,0.0
+kujiale_0240_32_1,kujiale_0240,1.5348839212894467,1,1.0,1.0
+kujiale_0240_48_1,kujiale_0240,0.8650982576265832,1,1.0,0.890123182847532
+kujiale_0240_40_1,kujiale_0240,1.7888093337210187,1,1.0,1.0
+kujiale_0240_55_0,kujiale_0240,3.1843635004282524,0,1.0,0.0
+kujiale_0240_6_3,kujiale_0240,1.3056400123087386,1,1.0,1.0
+kujiale_0240_16_3,kujiale_0240,2.6099634232898654,1,1.0,0.9566400753790925
+kujiale_0240_47_3,kujiale_0240,2.5661970084111014,1,1.0,0.6124109079713407
+kujiale_0240_9_0,kujiale_0240,4.943650570365787,0,0.0,0.0
+kujiale_0240_20_1,kujiale_0240,2.072177836054381,1,1.0,0.6865179208734616
+kujiale_0240_33_2,kujiale_0240,6.665330212827378,0,0.0,0.0
+kujiale_0234_46_3,kujiale_0234,7.301392117850446,0,0.0,0.0
+kujiale_0234_62_2,kujiale_0234,6.532767110186247,0,0.0,0.0
+kujiale_0234_54_0,kujiale_0234,0.9590706619714142,1,1.0,0.978314869526096
+kujiale_0234_59_2,kujiale_0234,2.559489585198155,1,1.0,1.0
+kujiale_0234_51_4,kujiale_0234,9.028762109482244,0,0.0,0.0
+kujiale_0234_24_3,kujiale_0234,7.85611458451816,0,1.0,0.0
+kujiale_0234_53_0,kujiale_0234,4.795437796798018,0,0.0,0.0
+kujiale_0234_38_3,kujiale_0234,1.8459325144526302,1,1.0,1.0
+kujiale_0234_43_3,kujiale_0234,7.446250991800351,0,0.0,0.0
+kujiale_0234_20_0,kujiale_0234,1.834864660675553,1,1.0,0.9575100024396577
+kujiale_0234_50_1,kujiale_0234,2.6076929337937376,1,1.0,0.9126374665228435
+kujiale_0234_45_0,kujiale_0234,11.103565362099085,0,0.0,0.0
+kujiale_0234_61_4,kujiale_0234,5.294329833538806,0,0.0,0.0
+kujiale_0234_52_0,kujiale_0234,7.593422777600212,0,0.0,0.0
+kujiale_0234_22_3,kujiale_0234,3.363116787113761,0,0.0,0.0
+kujiale_0234_32_4,kujiale_0234,9.183574846895198,0,0.0,0.0
+kujiale_0234_48_2,kujiale_0234,5.333220688101263,0,0.0,0.0
+kujiale_0234_30_2,kujiale_0234,1.8246342251018244,1,1.0,1.0
+kujiale_0234_5_4,kujiale_0234,0.8775235214394908,1,1.0,0.964511226563488
+kujiale_0234_6_2,kujiale_0234,6.0125980191568225,0,0.0,0.0
+kujiale_0234_29_1,kujiale_0234,8.800546181258497,0,1.0,0.0
+kujiale_0234_55_1,kujiale_0234,10.770236123262036,0,0.0,0.0

eval_results/shards/shard0_serial/server.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccd9d346b17d9b4eb3474f712129cc56ecbaa11f756498243cf54796e1bd8cd5
+size 73317

eval_results/shards/shard1_retry/eval.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b38eaa9e70e418c484dd752dcf92946d86ddacf03215f11fa539dbe75f02c3a
+size 6879174

eval_results/shards/shard1_retry/eval_cfg.py

@@ -0,0 +1,105 @@
+"""RyWorld eval cfg — VLNVerse coarse / val_unseen, H1 robot at 256x256.
+
+Forked from
+``third_party/vlnverse_emr/scripts/eval/configs/h1_cma_clip_cfg_vlnverse_coarse.py``
+with the minimum diffs needed for our model:
+
+  - ``agent.model_name='ryworld'`` so Agent.init dispatches to RyWorldAgent.
+  - ``split_data_types=['val_unseen']`` only (CMA cfg also includes val_seen).
+  - ``task_name`` carries our run identity for ``logs/<task>/`` separation.
+
+Everything else (camera 256x256, ``torso_link/h1_pano_camera_0``,
+``h1_vln_pointcloud.usd``, ``vln_pe`` env, ``flash`` mode, max_step=500,
+formal instruction) is kept identical to the CMA-CLIP baseline so our
+numbers go on the same table.
+
+Plugin registration: this cfg's top-level ``import
+ryworld.eval.vlnverse_plugin`` triggers ``@Agent.register('ryworld')`` as a
+side effect, which is what lets vlnverse_emr's dispatcher find our agent
+without a fork. ``importlib.util.spec_from_file_location`` (used by both
+``server.py`` and ``eval.py``) executes this module top-to-bottom, so by the
+time ``Agent.init`` runs the registration is already in place.
+"""
+
+from __future__ import annotations
+
+# ---------------------------------------------------------------------------
+# 1. Make the RyWorld package importable.
+#    vlnverse_emr's launchers run from inside ``third_party/vlnverse_emr``
+#    (their ``cwd``), so ``ryworld.*`` is not on ``sys.path`` by default. We
+#    walk up from this file: cfg -> configs -> vlnverse_plugin -> eval ->
+#    ryworld -> repo root.
+# ---------------------------------------------------------------------------
+import sys
+from pathlib import Path
+
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+# ---------------------------------------------------------------------------
+# 2. Import the plugin package — side effect: ``@Agent.register('ryworld')``.
+# ---------------------------------------------------------------------------
+import ryworld.eval.vlnverse_plugin  # noqa: E402, F401
+
+# ---------------------------------------------------------------------------
+# 3. Build the eval cfg using vlnverse_emr's pydantic config classes.
+# ---------------------------------------------------------------------------
+from vlnverse.configs.agent import AgentCfg  # type: ignore[import-not-found]  # noqa: E402
+from vlnverse.configs.evaluator import (  # type: ignore[import-not-found]  # noqa: E402
+    EnvCfg,
+    EvalCfg,
+    EvalDatasetCfg,
+    SceneCfg,
+    TaskCfg,
+)
+
+eval_cfg = EvalCfg(
+    agent=AgentCfg(
+        # Server port — keep clear of CMA(8087)/N1(8028) so we can run side-by-side.
+        server_port=8092,
+        model_name="ryworld",
+        # Dummy skeleton agent ignores ckpt_path; real wrapper will load this.
+        ckpt_path="/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+        model_settings={
+            "stop_threshold_override": 0.95,
+            "train_config_path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/stage1_discrete.yaml",
+            "train_config_overrides": ["/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/overlays/a100_4gpu_discrete.yaml"],
+        },
+    ),
+    env=EnvCfg(
+        env_type="vln_pe",
+        env_settings={
+            "use_fabric": False,
+            "headless": True,
+        },
+    ),
+    task=TaskCfg(
+        task_name="20260513T132258_step0015000_stop0.95_full835_thr095_shard1_retry",
+        task_settings={
+            "env_num": 1,
+            "use_distributed": False,
+            "proc_num": 1,
+            "max_step": 500,
+            "warm_up_step": 500,
+        },
+        scene=SceneCfg(
+            scene_type="kujiale",
+            scene_data_dir="data/scene_data/vlnverse",
+        ),
+        robot_name="h1",
+        robot_flash=True,
+        robot_usd_path="data/Embodiments/vln-pe/h1/h1_vln_pointcloud.usd",
+        camera_resolution=[256, 256],
+        camera_prim_path="torso_link/h1_pano_camera_0",
+    ),
+    dataset=EvalDatasetCfg(
+        dataset_type="kujiale",
+        dataset_settings={
+            "base_data_dir": "data/vlnverse/raw_data/vlnverse/coarse_shard1",
+            "split_data_types": ["val_unseen"],
+            "filter_stairs": False,
+            "instruction_type": "formal",
+        },
+    ),
+)

eval_results/shards/shard1_retry/eval_result.log

@@ -0,0 +1,16 @@
+[split:val_unseen] total 104 data
+############[val_unseen]#############
+TL = 727.2584597270511 / 104 = 6.9929
+NE = 429.5710585891784 / 104 = 4.1305
+FR = 0 / 104 = 0.0%
+StR = 0 / 104 = 0%
+OS = 62.0 / 104 = 59.62%
+SR = 47.0 / 104 = 45.190000000000005%
+SPL = 45.52605968207551 / 104 = 43.78%
+detail:
+[reach_goal]:47
+[not_reach_goal]:54
+[success]:46
+[exceed_total_max_step]:4
+[fall]:0
+##########################

eval_results/shards/shard1_retry/meta.json

@@ -0,0 +1,162 @@
+{
+  "eval_id": "20260513T132258_step0015000_stop0.95_full835_thr095_shard1_retry",
+  "tag": "full835_thr095_shard1_retry",
+  "timestamp_recorded": "2026-05-13T18:03:06+0800",
+  "ckpt": {
+    "path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+    "step": 15000,
+    "size_bytes": 2812658821
+  },
+  "code_state": {
+    "git_hash": "a4fa65370358d6f251e158f03aced7d30f167da6",
+    "branch": "feat/batched-trunk-forward",
+    "dirty": true,
+    "dirty_files": [
+      "docs/eval_ledger.jsonl",
+      "run_train.sh",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD0_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD1_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD2_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD3_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD4_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD5_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD6_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD7_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu0_thr091_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu1_thr092_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu2_thr093_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu3_thr094_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu4_thr095_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu5_thr096_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu6_thr097_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu7_thr098_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v7_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v8_port8092.py",
+      "scripts/eval/run_eval_structured.sh",
+      "third_party/vlnverse_emr",
+      "yworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen.py"
+    ],
+    "dirty_patch_md5": "aa7101e2b529b2f5596267cd1e8073ee"
+  },
+  "eval_config": {
+    "framework": "vlnverse_emr",
+    "base_cfg": "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD1_port8092.py",
+    "derived_cfg": "checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/evals/20260513T132258_step0015000_stop0.95_full835_thr095_shard1_retry/eval_cfg.py",
+    "stop_threshold": 0.95,
+    "task_name": "20260513T132258_step0015000_stop0.95_full835_thr095_shard1_retry"
+  },
+  "data_state": {
+    "scenes_used": [
+      "kujiale_0113",
+      "kujiale_0123",
+      "kujiale_0159",
+      "kujiale_0165",
+      "kujiale_0240"
+    ],
+    "scenes_with_hdr": [
+      "kujiale_0113",
+      "kujiale_0159",
+      "kujiale_0165",
+      "kujiale_0240"
+    ],
+    "scenes_missing_hdr": [
+      "kujiale_0123"
+    ]
+  },
+  "hardware": {
+    "host": "di-20260502152021-n5hs7",
+    "gpus": [
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      }
+    ],
+    "count": 8
+  },
+  "results_aggregate": {
+    "TL": 6.9929,
+    "NE": 4.1305,
+    "FR": 0.0,
+    "StR": 0.0,
+    "OS": 0.5962,
+    "SR": 0.4519,
+    "SPL": 0.4378,
+    "n_reach_goal": 47,
+    "n_not_reach_goal": 54,
+    "n_success": 46,
+    "n_fall": 0,
+    "n_total": 104
+  },
+  "results_per_scene": {
+    "kujiale_0240": {
+      "n": 11,
+      "SR": 0.6363636363636364,
+      "avg_NE": 2.49024044625038,
+      "avg_SPL": 0.6282146491471609
+    },
+    "kujiale_0165": {
+      "n": 18,
+      "SR": 0.3888888888888889,
+      "avg_NE": 4.661110946215886,
+      "avg_SPL": 0.3429450023641647
+    },
+    "kujiale_0159": {
+      "n": 25,
+      "SR": 0.64,
+      "avg_NE": 2.89723540705876,
+      "avg_SPL": 0.638133977009026
+    },
+    "kujiale_0123": {
+      "n": 25,
+      "SR": 0.48,
+      "avg_NE": 4.399718945699857,
+      "avg_SPL": 0.46431919367843577
+    },
+    "kujiale_0113": {
+      "n": 24,
+      "SR": 0.20833333333333334,
+      "avg_NE": 5.443181415820793,
+      "avg_SPL": 0.2033899679881351
+    }
+  },
+  "stop_head_stats": {
+    "n_samples": 24240,
+    "stop_prob_min": 0.345,
+    "stop_prob_p50": 0.786,
+    "stop_prob_p90": 0.904,
+    "stop_prob_max": 0.976,
+    "stop_prob_mean": 0.7602361386138585,
+    "fire_rate_pathA": 0.02277227722772277,
+    "fire_rate_pathB": 0.004909240924092409,
+    "no_stop_rate": 0.9723184818481848
+  },
+  "n_episodes_recorded": 103
+}

eval_results/shards/shard1_retry/per_episode.csv

@@ -0,0 +1,104 @@
+episode_id,scene,NE,success,osr,SPL
+kujiale_0240_62_4,kujiale_0240,4.942261613479123,0,1.0,0.0
+kujiale_0240_51_4,kujiale_0240,3.6418222074763476,0,0.0,0.0
+kujiale_0240_10_0,kujiale_0240,0.9040651909662726,1,1.0,1.0
+kujiale_0240_43_3,kujiale_0240,0.6236469389572344,1,1.0,1.0
+kujiale_0240_59_2,kujiale_0240,1.5700042360255968,1,1.0,0.9103611406187703
+kujiale_0240_4_4,kujiale_0240,1.298973825626585,1,1.0,1.0
+kujiale_0240_26_1,kujiale_0240,1.5897949875278299,1,1.0,1.0
+kujiale_0240_38_4,kujiale_0240,3.445574887117299,0,1.0,0.0
+kujiale_0240_29_4,kujiale_0240,5.710757452121341,0,0.0,0.0
+kujiale_0240_58_3,kujiale_0240,1.199786330887976,1,1.0,1.0
+kujiale_0240_22_3,kujiale_0240,2.465957238568579,1,1.0,1.0
+kujiale_0165_23_1,kujiale_0165,6.027670354089401,0,0.0,0.0
+kujiale_0165_62_2,kujiale_0165,12.666122873398056,0,0.0,0.0
+kujiale_0165_24_3,kujiale_0165,4.977560749465032,0,1.0,0.0
+kujiale_0165_7_4,kujiale_0165,1.0912870921871352,1,1.0,0.9806316808011822
+kujiale_0165_10_0,kujiale_0165,3.752763098395252,0,0.0,0.0
+kujiale_0165_54_0,kujiale_0165,3.431245147923428,0,1.0,0.0
+kujiale_0165_12_3,kujiale_0165,4.6006677333205745,0,1.0,0.0
+kujiale_0165_21_2,kujiale_0165,5.662265748606484,0,1.0,0.0
+kujiale_0165_33_2,kujiale_0165,9.188082307071419,0,1.0,0.0
+kujiale_0165_38_3,kujiale_0165,6.9639280639995835,0,0.0,0.0
+kujiale_0165_5_4,kujiale_0165,2.69722042837169,1,1.0,0.7682037876812625
+kujiale_0165_22_3,kujiale_0165,1.0936031902595662,1,1.0,1.0
+kujiale_0165_56_3,kujiale_0165,2.4900383940250386,1,1.0,0.7860083859039844
+kujiale_0165_47_1,kujiale_0165,2.2944095712192345,1,1.0,0.718505763686361
+kujiale_0165_8_3,kujiale_0165,1.8044616544239864,1,1.0,0.919660424482175
+kujiale_0165_32_4,kujiale_0165,2.4500719280490664,1,1.0,1.0
+kujiale_0165_43_3,kujiale_0165,7.552867746564375,0,0.0,0.0
+kujiale_0165_64_0,kujiale_0165,5.155730950516618,0,0.0,0.0
+kujiale_0159_32_2,kujiale_0159,0.16724083521990998,1,1.0,1.0
+kujiale_0159_10_2,kujiale_0159,1.0776604262024523,1,1.0,1.0
+kujiale_0159_21_2,kujiale_0159,1.0969534804605716,1,1.0,1.0
+kujiale_0159_20_3,kujiale_0159,1.0226992755936866,1,1.0,1.0
+kujiale_0159_27_1,kujiale_0159,2.822907934592298,1,1.0,1.0
+kujiale_0159_2_2,kujiale_0159,8.553785123289744,0,0.0,0.0
+kujiale_0159_4_1,kujiale_0159,0.6489613398343561,1,1.0,1.0
+kujiale_0159_5_2,kujiale_0159,1.5783941621824162,1,1.0,1.0
+kujiale_0159_30_3,kujiale_0159,5.691582680229409,0,0.0,0.0
+kujiale_0159_0_0,kujiale_0159,1.9310091653560835,1,1.0,1.0
+kujiale_0159_23_3,kujiale_0159,1.1441196740775599,1,1.0,1.0
+kujiale_0159_58_0,kujiale_0159,6.045398174003273,0,0.0,0.0
+kujiale_0159_61_0,kujiale_0159,3.4147656981548455,0,1.0,0.0
+kujiale_0159_59_1,kujiale_0159,6.185459040437604,0,0.0,0.0
+kujiale_0159_52_0,kujiale_0159,1.965610897848253,1,1.0,1.0
+kujiale_0159_42_4,kujiale_0159,0.6459695585550327,1,1.0,0.9533494252256478
+kujiale_0159_46_4,kujiale_0159,2.108194707553545,1,1.0,1.0
+kujiale_0159_54_4,kujiale_0159,1.585878801052509,1,1.0,1.0
+kujiale_0159_55_1,kujiale_0159,3.7101352239155347,0,1.0,0.0
+kujiale_0159_60_2,kujiale_0159,6.587826835435354,0,1.0,0.0
+kujiale_0159_22_4,kujiale_0159,1.2883233116001325,1,1.0,1.0
+kujiale_0159_1_0,kujiale_0159,5.2349019321311845,0,0.0,0.0
+kujiale_0159_53_0,kujiale_0159,0.7466049194588325,1,1.0,1.0
+kujiale_0159_8_0,kujiale_0159,2.1563831406999365,1,1.0,1.0
+kujiale_0159_26_4,kujiale_0159,5.020118838584473,0,0.0,0.0
+kujiale_0123_fix_40_0,kujiale_0123,9.181934548263152,0,0.0,0.0
+kujiale_0123_fix_18_3,kujiale_0123,2.9198449515195155,1,1.0,1.0
+kujiale_0123_fix_44_2,kujiale_0123,0.4834924277180767,1,1.0,1.0
+kujiale_0123_fix_8_2,kujiale_0123,2.5632601558512196,1,1.0,1.0
+kujiale_0123_fix_35_1,kujiale_0123,8.191500737381752,0,0.0,0.0
+kujiale_0123_fix_9_2,kujiale_0123,4.812722939332428,0,1.0,0.0
+kujiale_0123_fix_28_0,kujiale_0123,4.2753365449543095,0,1.0,0.0
+kujiale_0123_fix_10_3,kujiale_0123,2.4914601886240937,1,1.0,1.0
+kujiale_0123_fix_14_0,kujiale_0123,1.5147510068804906,1,1.0,1.0
+kujiale_0123_fix_27_0,kujiale_0123,1.8584728186391275,1,1.0,0.853408214013301
+kujiale_0123_fix_41_1,kujiale_0123,5.568056406840726,0,0.0,0.0
+kujiale_0123_fix_2_3,kujiale_0123,7.078084526730988,0,0.0,0.0
+kujiale_0123_fix_37_3,kujiale_0123,3.0783415118939876,0,0.0,0.0
+kujiale_0123_fix_31_0,kujiale_0123,1.3509477850893081,1,1.0,1.0
+kujiale_0123_fix_20_4,kujiale_0123,7.068725069465462,0,0.0,0.0
+kujiale_0123_fix_35_4,kujiale_0123,1.782784775671893,1,1.0,1.0
+kujiale_0123_fix_36_1,kujiale_0123,1.0084550667965,1,1.0,1.0
+kujiale_0123_fix_43_0,kujiale_0123,9.7564065079623,0,0.0,0.0
+kujiale_0123_fix_39_1,kujiale_0123,3.7762138198586648,0,1.0,0.0
+kujiale_0123_fix_16_1,kujiale_0123,9.274062711501674,0,0.0,0.0
+kujiale_0123_fix_38_2,kujiale_0123,7.31442752623403,0,0.0,0.0
+kujiale_0123_fix_12_1,kujiale_0123,2.8491269192841027,1,1.0,1.0
+kujiale_0123_fix_21_2,kujiale_0123,1.9043952526074752,1,1.0,0.754571627947592
+kujiale_0123_fix_1_2,kujiale_0123,0.1033247402329383,1,1.0,1.0
+kujiale_0123_fix_3_2,kujiale_0123,9.786844703162231,0,0.0,0.0
+kujiale_0113_67_0,kujiale_0113,5.2088164699179185,0,0.0,0.0
+kujiale_0113_57_1,kujiale_0113,1.6303966621307207,1,1.0,1.0
+kujiale_0113_13_3,kujiale_0113,1.7552671120403576,1,1.0,1.0
+kujiale_0113_61_0,kujiale_0113,1.4879522905298834,1,1.0,1.0
+kujiale_0113_49_0,kujiale_0113,6.412942297278781,0,0.0,0.0
+kujiale_0113_17_0,kujiale_0113,9.15748732515745,0,0.0,0.0
+kujiale_0113_91_3,kujiale_0113,7.91954457740553,0,0.0,0.0
+kujiale_0113_87_2,kujiale_0113,4.167005777552314,0,0.0,0.0
+kujiale_0113_94_4,kujiale_0113,1.52558866333996,1,1.0,0.8813592317152429
+kujiale_0113_5_2,kujiale_0113,3.963741242948207,0,0.0,0.0
+kujiale_0113_44_1,kujiale_0113,5.58432344478236,0,0.0,0.0
+kujiale_0113_86_0,kujiale_0113,6.653558391272165,0,0.0,0.0
+kujiale_0113_4_2,kujiale_0113,5.519723777740512,0,0.0,0.0
+kujiale_0113_88_4,kujiale_0113,3.2242844006650744,0,1.0,0.0
+kujiale_0113_38_0,kujiale_0113,3.17432351022298,0,0.0,0.0
+kujiale_0113_43_3,kujiale_0113,3.2150738338318328,0,0.0,0.0
+kujiale_0113_11_4,kujiale_0113,10.497963353109716,0,0.0,0.0
+kujiale_0113_58_4,kujiale_0113,11.103238190126685,0,0.0,0.0
+kujiale_0113_2_3,kujiale_0113,8.666349719229682,0,0.0,0.0
+kujiale_0113_42_0,kujiale_0113,3.2974669758424153,0,0.0,0.0
+kujiale_0113_65_3,kujiale_0113,6.055023790924712,0,1.0,0.0
+kujiale_0113_75_1,kujiale_0113,1.7923248899274749,1,1.0,1.0
+kujiale_0113_23_2,kujiale_0113,7.816515135351559,0,0.0,0.0
+kujiale_0113_71_0,kujiale_0113,10.807442148370743,0,0.0,0.0

eval_results/shards/shard1_retry/server.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f899de72a338186095da8c446d80b0131abcc91045e52b83a466b57629eb5d7b
+size 104442

eval_results/shards/shard1_serial/eval_cfg.py

@@ -0,0 +1,105 @@
+"""RyWorld eval cfg — VLNVerse coarse / val_unseen, H1 robot at 256x256.
+
+Forked from
+``third_party/vlnverse_emr/scripts/eval/configs/h1_cma_clip_cfg_vlnverse_coarse.py``
+with the minimum diffs needed for our model:
+
+  - ``agent.model_name='ryworld'`` so Agent.init dispatches to RyWorldAgent.
+  - ``split_data_types=['val_unseen']`` only (CMA cfg also includes val_seen).
+  - ``task_name`` carries our run identity for ``logs/<task>/`` separation.
+
+Everything else (camera 256x256, ``torso_link/h1_pano_camera_0``,
+``h1_vln_pointcloud.usd``, ``vln_pe`` env, ``flash`` mode, max_step=500,
+formal instruction) is kept identical to the CMA-CLIP baseline so our
+numbers go on the same table.
+
+Plugin registration: this cfg's top-level ``import
+ryworld.eval.vlnverse_plugin`` triggers ``@Agent.register('ryworld')`` as a
+side effect, which is what lets vlnverse_emr's dispatcher find our agent
+without a fork. ``importlib.util.spec_from_file_location`` (used by both
+``server.py`` and ``eval.py``) executes this module top-to-bottom, so by the
+time ``Agent.init`` runs the registration is already in place.
+"""
+
+from __future__ import annotations
+
+# ---------------------------------------------------------------------------
+# 1. Make the RyWorld package importable.
+#    vlnverse_emr's launchers run from inside ``third_party/vlnverse_emr``
+#    (their ``cwd``), so ``ryworld.*`` is not on ``sys.path`` by default. We
+#    walk up from this file: cfg -> configs -> vlnverse_plugin -> eval ->
+#    ryworld -> repo root.
+# ---------------------------------------------------------------------------
+import sys
+from pathlib import Path
+
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+# ---------------------------------------------------------------------------
+# 2. Import the plugin package — side effect: ``@Agent.register('ryworld')``.
+# ---------------------------------------------------------------------------
+import ryworld.eval.vlnverse_plugin  # noqa: E402, F401
+
+# ---------------------------------------------------------------------------
+# 3. Build the eval cfg using vlnverse_emr's pydantic config classes.
+# ---------------------------------------------------------------------------
+from vlnverse.configs.agent import AgentCfg  # type: ignore[import-not-found]  # noqa: E402
+from vlnverse.configs.evaluator import (  # type: ignore[import-not-found]  # noqa: E402
+    EnvCfg,
+    EvalCfg,
+    EvalDatasetCfg,
+    SceneCfg,
+    TaskCfg,
+)
+
+eval_cfg = EvalCfg(
+    agent=AgentCfg(
+        # Server port — keep clear of CMA(8087)/N1(8028) so we can run side-by-side.
+        server_port=8092,
+        model_name="ryworld",
+        # Dummy skeleton agent ignores ckpt_path; real wrapper will load this.
+        ckpt_path="/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+        model_settings={
+            "stop_threshold_override": 0.95,
+            "train_config_path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/stage1_discrete.yaml",
+            "train_config_overrides": ["/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/overlays/a100_4gpu_discrete.yaml"],
+        },
+    ),
+    env=EnvCfg(
+        env_type="vln_pe",
+        env_settings={
+            "use_fabric": False,
+            "headless": True,
+        },
+    ),
+    task=TaskCfg(
+        task_name="20260513T032430_step0015000_stop0.95_full835_thr095_shard1_serial",
+        task_settings={
+            "env_num": 1,
+            "use_distributed": False,
+            "proc_num": 1,
+            "max_step": 500,
+            "warm_up_step": 500,
+        },
+        scene=SceneCfg(
+            scene_type="kujiale",
+            scene_data_dir="data/scene_data/vlnverse",
+        ),
+        robot_name="h1",
+        robot_flash=True,
+        robot_usd_path="data/Embodiments/vln-pe/h1/h1_vln_pointcloud.usd",
+        camera_resolution=[256, 256],
+        camera_prim_path="torso_link/h1_pano_camera_0",
+    ),
+    dataset=EvalDatasetCfg(
+        dataset_type="kujiale",
+        dataset_settings={
+            "base_data_dir": "data/vlnverse/raw_data/vlnverse/coarse_shard1",
+            "split_data_types": ["val_unseen"],
+            "filter_stairs": False,
+            "instruction_type": "formal",
+        },
+    ),
+)

eval_results/shards/shard2_serial/eval.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:664a13985d7dd8ee8296dd5d15fc369718fa1f85deab48af57c3e70cf6bfc354
+size 6218227

eval_results/shards/shard2_serial/eval_cfg.py

@@ -0,0 +1,105 @@
+"""RyWorld eval cfg — VLNVerse coarse / val_unseen, H1 robot at 256x256.
+
+Forked from
+``third_party/vlnverse_emr/scripts/eval/configs/h1_cma_clip_cfg_vlnverse_coarse.py``
+with the minimum diffs needed for our model:
+
+  - ``agent.model_name='ryworld'`` so Agent.init dispatches to RyWorldAgent.
+  - ``split_data_types=['val_unseen']`` only (CMA cfg also includes val_seen).
+  - ``task_name`` carries our run identity for ``logs/<task>/`` separation.
+
+Everything else (camera 256x256, ``torso_link/h1_pano_camera_0``,
+``h1_vln_pointcloud.usd``, ``vln_pe`` env, ``flash`` mode, max_step=500,
+formal instruction) is kept identical to the CMA-CLIP baseline so our
+numbers go on the same table.
+
+Plugin registration: this cfg's top-level ``import
+ryworld.eval.vlnverse_plugin`` triggers ``@Agent.register('ryworld')`` as a
+side effect, which is what lets vlnverse_emr's dispatcher find our agent
+without a fork. ``importlib.util.spec_from_file_location`` (used by both
+``server.py`` and ``eval.py``) executes this module top-to-bottom, so by the
+time ``Agent.init`` runs the registration is already in place.
+"""
+
+from __future__ import annotations
+
+# ---------------------------------------------------------------------------
+# 1. Make the RyWorld package importable.
+#    vlnverse_emr's launchers run from inside ``third_party/vlnverse_emr``
+#    (their ``cwd``), so ``ryworld.*`` is not on ``sys.path`` by default. We
+#    walk up from this file: cfg -> configs -> vlnverse_plugin -> eval ->
+#    ryworld -> repo root.
+# ---------------------------------------------------------------------------
+import sys
+from pathlib import Path
+
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+# ---------------------------------------------------------------------------
+# 2. Import the plugin package — side effect: ``@Agent.register('ryworld')``.
+# ---------------------------------------------------------------------------
+import ryworld.eval.vlnverse_plugin  # noqa: E402, F401
+
+# ---------------------------------------------------------------------------
+# 3. Build the eval cfg using vlnverse_emr's pydantic config classes.
+# ---------------------------------------------------------------------------
+from vlnverse.configs.agent import AgentCfg  # type: ignore[import-not-found]  # noqa: E402
+from vlnverse.configs.evaluator import (  # type: ignore[import-not-found]  # noqa: E402
+    EnvCfg,
+    EvalCfg,
+    EvalDatasetCfg,
+    SceneCfg,
+    TaskCfg,
+)
+
+eval_cfg = EvalCfg(
+    agent=AgentCfg(
+        # Server port — keep clear of CMA(8087)/N1(8028) so we can run side-by-side.
+        server_port=8093,
+        model_name="ryworld",
+        # Dummy skeleton agent ignores ckpt_path; real wrapper will load this.
+        ckpt_path="/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+        model_settings={
+            "stop_threshold_override": 0.95,
+            "train_config_path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/stage1_discrete.yaml",
+            "train_config_overrides": ["/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/overlays/a100_4gpu_discrete.yaml"],
+        },
+    ),
+    env=EnvCfg(
+        env_type="vln_pe",
+        env_settings={
+            "use_fabric": False,
+            "headless": True,
+        },
+    ),
+    task=TaskCfg(
+        task_name="20260513T032530_step0015000_stop0.95_full835_thr095_shard2_serial",
+        task_settings={
+            "env_num": 1,
+            "use_distributed": False,
+            "proc_num": 1,
+            "max_step": 500,
+            "warm_up_step": 500,
+        },
+        scene=SceneCfg(
+            scene_type="kujiale",
+            scene_data_dir="data/scene_data/vlnverse",
+        ),
+        robot_name="h1",
+        robot_flash=True,
+        robot_usd_path="data/Embodiments/vln-pe/h1/h1_vln_pointcloud.usd",
+        camera_resolution=[256, 256],
+        camera_prim_path="torso_link/h1_pano_camera_0",
+    ),
+    dataset=EvalDatasetCfg(
+        dataset_type="kujiale",
+        dataset_settings={
+            "base_data_dir": "data/vlnverse/raw_data/vlnverse/coarse_shard2",
+            "split_data_types": ["val_unseen"],
+            "filter_stairs": False,
+            "instruction_type": "formal",
+        },
+    ),
+)

eval_results/shards/shard2_serial/eval_result.log

@@ -0,0 +1,16 @@
+[split:val_unseen] total 104 data
+############[val_unseen]#############
+TL = 641.8399859772071 / 104 = 6.1715
+NE = 375.5218178826217 / 104 = 3.6108
+FR = 1 / 104 = 0.96%
+StR = 0 / 104 = 0%
+OS = 65.0 / 104 = 62.5%
+SR = 56.0 / 104 = 53.849999999999994%
+SPL = 54.239708538828175 / 104 = 52.15%
+detail:
+[reach_goal]:56
+[not_reach_goal]:47
+[success]:55
+[exceed_total_max_step]:1
+[fall]:1
+##########################

eval_results/shards/shard2_serial/meta.json

@@ -0,0 +1,162 @@
+{
+  "eval_id": "20260513T032530_step0015000_stop0.95_full835_thr095_shard2_serial",
+  "tag": "full835_thr095_shard2_serial",
+  "timestamp_recorded": "2026-05-13T16:23:15+0800",
+  "ckpt": {
+    "path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+    "step": 15000,
+    "size_bytes": 2812658821
+  },
+  "code_state": {
+    "git_hash": "a4fa65370358d6f251e158f03aced7d30f167da6",
+    "branch": "feat/batched-trunk-forward",
+    "dirty": true,
+    "dirty_files": [
+      "docs/eval_ledger.jsonl",
+      "run_train.sh",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD0_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD1_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD2_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD3_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD4_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD5_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD6_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD7_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu0_thr091_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu1_thr092_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu2_thr093_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu3_thr094_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu4_thr095_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu5_thr096_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu6_thr097_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu7_thr098_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v7_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v8_port8092.py",
+      "scripts/eval/run_eval_structured.sh",
+      "third_party/vlnverse_emr",
+      "yworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen.py"
+    ],
+    "dirty_patch_md5": "aa7101e2b529b2f5596267cd1e8073ee"
+  },
+  "eval_config": {
+    "framework": "vlnverse_emr",
+    "base_cfg": "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD2_port8093.py",
+    "derived_cfg": "checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/evals/20260513T032530_step0015000_stop0.95_full835_thr095_shard2_serial/eval_cfg.py",
+    "stop_threshold": 0.95,
+    "task_name": "20260513T032530_step0015000_stop0.95_full835_thr095_shard2_serial"
+  },
+  "data_state": {
+    "scenes_used": [
+      "kujiale_0128",
+      "kujiale_0165",
+      "kujiale_0257",
+      "kujiale_0259",
+      "kujiale_0285"
+    ],
+    "scenes_with_hdr": [
+      "kujiale_0165",
+      "kujiale_0257",
+      "kujiale_0259",
+      "kujiale_0285"
+    ],
+    "scenes_missing_hdr": [
+      "kujiale_0128"
+    ]
+  },
+  "hardware": {
+    "host": "di-20260502152021-n5hs7",
+    "gpus": [
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      }
+    ],
+    "count": 8
+  },
+  "results_aggregate": {
+    "TL": 6.1715,
+    "NE": 3.6108,
+    "FR": 0.0096,
+    "StR": 0.0,
+    "OS": 0.625,
+    "SR": 0.5385,
+    "SPL": 0.5215,
+    "n_reach_goal": 56,
+    "n_not_reach_goal": 47,
+    "n_success": 55,
+    "n_fall": 1,
+    "n_total": 104
+  },
+  "results_per_scene": {
+    "kujiale_0285": {
+      "n": 25,
+      "SR": 0.72,
+      "avg_NE": 2.5282313948518076,
+      "avg_SPL": 0.7002948465313784
+    },
+    "kujiale_0259": {
+      "n": 24,
+      "SR": 0.5416666666666666,
+      "avg_NE": 4.19641644940283,
+      "avg_SPL": 0.5221671770667976
+    },
+    "kujiale_0257": {
+      "n": 22,
+      "SR": 0.36363636363636365,
+      "avg_NE": 4.64986981016304,
+      "avg_SPL": 0.3437024146320122
+    },
+    "kujiale_0165": {
+      "n": 7,
+      "SR": 0.2857142857142857,
+      "avg_NE": 4.433341278678205,
+      "avg_SPL": 0.2376640031541409
+    },
+    "kujiale_0128": {
+      "n": 25,
+      "SR": 0.56,
+      "avg_NE": 3.072718249976478,
+      "avg_SPL": 0.56
+    }
+  },
+  "stop_head_stats": {
+    "n_samples": 16412,
+    "stop_prob_min": 0.333,
+    "stop_prob_p50": 0.789,
+    "stop_prob_p90": 0.906,
+    "stop_prob_max": 0.976,
+    "stop_prob_mean": 0.7804933585181553,
+    "fire_rate_pathA": 0.028820375335120642,
+    "fire_rate_pathB": 0.007677309285888374,
+    "no_stop_rate": 0.963502315378991
+  },
+  "n_episodes_recorded": 103
+}

eval_results/shards/shard2_serial/per_episode.csv

@@ -0,0 +1,104 @@
+episode_id,scene,NE,success,osr,SPL
+kujiale_0285_1_3,kujiale_0285,2.3909762412297613,1,1.0,0.921107344565759
+kujiale_0285_8_1,kujiale_0285,5.249190769617014,0,0.0,0.0
+kujiale_0285_10_3,kujiale_0285,2.2752195084946116,1,1.0,0.7362517957868427
+kujiale_0285_0_1,kujiale_0285,1.9830840957271412,1,1.0,1.0
+kujiale_0285_3_3,kujiale_0285,2.224111711486303,1,1.0,1.0
+kujiale_0285_9_4,kujiale_0285,0.17370096355269726,1,1.0,1.0
+kujiale_0285_5_4,kujiale_0285,4.903995451059142,0,0.0,0.0
+kujiale_0285_4_1,kujiale_0285,1.6509242011403942,1,1.0,1.0
+kujiale_0285_12_2,kujiale_0285,1.770056203347089,1,1.0,1.0
+kujiale_0285_9_3,kujiale_0285,2.65500808649692,1,1.0,1.0
+kujiale_0285_4_4,kujiale_0285,0.724436486847339,1,1.0,1.0
+kujiale_0285_6_3,kujiale_0285,1.0609340762156927,1,1.0,1.0
+kujiale_0285_11_4,kujiale_0285,2.157500237993809,1,1.0,1.0
+kujiale_0285_3_2,kujiale_0285,3.0521264447016034,0,1.0,0.0
+kujiale_0285_14_0,kujiale_0285,1.0417544622336283,1,1.0,1.0
+kujiale_0285_0_4,kujiale_0285,3.0657326869417654,0,0.0,0.0
+kujiale_0285_14_4,kujiale_0285,1.4342102184153243,1,1.0,0.9369802531323966
+kujiale_0285_9_2,kujiale_0285,1.66571643003126,1,1.0,1.0
+kujiale_0285_12_1,kujiale_0285,1.2120685526662691,1,1.0,1.0
+kujiale_0285_5_2,kujiale_0285,1.92033632094134,1,1.0,1.0
+kujiale_0285_9_0,kujiale_0285,2.0439807577320854,1,1.0,1.0
+kujiale_0285_4_3,kujiale_0285,1.1500745428175634,1,1.0,0.9130317697994607
+kujiale_0285_6_4,kujiale_0285,4.835799208946911,0,0.0,0.0
+kujiale_0285_7_2,kujiale_0285,5.764229656282405,0,0.0,0.0
+kujiale_0285_13_3,kujiale_0285,6.800617556377135,0,0.0,0.0
+kujiale_0259_44_3,kujiale_0259,3.3932076402827245,0,0.0,0.0
+kujiale_0259_39_1,kujiale_0259,10.887980954236863,0,0.0,0.0
+kujiale_0259_22_3,kujiale_0259,1.0153972926542105,1,1.0,1.0
+kujiale_0259_51_3,kujiale_0259,11.000823181738827,0,0.0,0.0
+kujiale_0259_50_3,kujiale_0259,1.4984282969562441,1,1.0,0.8933697206531033
+kujiale_0259_37_4,kujiale_0259,7.657369587688456,0,0.0,0.0
+kujiale_0259_5_3,kujiale_0259,1.4865422661516816,1,1.0,0.9494898031680564
+kujiale_0259_4_3,kujiale_0259,10.563085255601342,0,0.0,0.0
+kujiale_0259_23_2,kujiale_0259,0.4323300532995481,1,1.0,1.0
+kujiale_0259_6_1,kujiale_0259,9.735610244425757,0,0.0,0.0
+kujiale_0259_30_4,kujiale_0259,1.1067864983927969,1,1.0,1.0
+kujiale_0259_35_0,kujiale_0259,1.7828874999186388,1,1.0,0.7680510702729438
+kujiale_0259_43_2,kujiale_0259,7.296450515476324,0,1.0,0.0
+kujiale_0259_47_1,kujiale_0259,9.62995167557212,0,0.0,0.0
+kujiale_0259_59_4,kujiale_0259,0.7856883910449153,1,1.0,1.0
+kujiale_0259_24_0,kujiale_0259,2.7711583508988467,1,1.0,1.0
+kujiale_0259_8_4,kujiale_0259,6.019685633128217,0,0.0,0.0
+kujiale_0259_49_3,kujiale_0259,4.717278219800516,0,1.0,0.0
+kujiale_0259_33_0,kujiale_0259,1.0084980912937713,1,1.0,1.0
+kujiale_0259_27_1,kujiale_0259,5.019541622588333,0,0.0,0.0
+kujiale_0259_34_0,kujiale_0259,1.219567958412943,1,1.0,1.0
+kujiale_0259_36_2,kujiale_0259,0.37452652011714727,1,1.0,1.0
+kujiale_0259_2_4,kujiale_0259,0.6105342808875904,1,1.0,1.0
+kujiale_0259_60_0,kujiale_0259,0.700664755100086,1,1.0,0.9211016555090371
+kujiale_0257_37_0,kujiale_0257,5.585119556173671,0,0.0,0.0
+kujiale_0257_67_4,kujiale_0257,1.9620655798145694,1,1.0,0.8782561480111803
+kujiale_0257_24_3,kujiale_0257,2.7126416107705595,1,1.0,0.758551740781148
+kujiale_0257_5_3,kujiale_0257,2.843321728794633,1,1.0,1.0
+kujiale_0257_66_2,kujiale_0257,1.057556355046107,1,1.0,1.0
+kujiale_0257_56_4,kujiale_0257,5.085173323409342,0,0.0,0.0
+kujiale_0257_72_2,kujiale_0257,6.525959974511102,0,0.0,0.0
+kujiale_0257_79_0,kujiale_0257,8.612904157099848,0,0.0,0.0
+kujiale_0257_29_3,kujiale_0257,7.342270144399747,0,0.0,0.0
+kujiale_0257_23_4,kujiale_0257,0.5616055851630964,1,1.0,1.0
+kujiale_0257_17_3,kujiale_0257,7.341487598769281,0,0.0,0.0
+kujiale_0257_49_1,kujiale_0257,10.009710650519624,0,1.0,0.0
+kujiale_0257_2_0,kujiale_0257,7.762460347753351,0,1.0,0.0
+kujiale_0257_21_0,kujiale_0257,1.505737258355273,1,1.0,1.0
+kujiale_0257_68_2,kujiale_0257,4.251789589473854,0,0.0,0.0
+kujiale_0257_84_0,kujiale_0257,3.6469352422274626,0,1.0,0.0
+kujiale_0257_46_4,kujiale_0257,6.646865334336234,0,0.0,0.0
+kujiale_0257_12_3,kujiale_0257,3.7097476693143383,0,1.0,0.0
+kujiale_0257_55_1,kujiale_0257,3.1538999569251143,0,0.0,0.0
+kujiale_0257_60_0,kujiale_0257,6.730909013105051,0,0.0,0.0
+kujiale_0257_1_1,kujiale_0257,2.5517785616217123,1,1.0,1.0
+kujiale_0257_69_1,kujiale_0257,2.6971965860028955,1,1.0,0.9246452331119398
+kujiale_0165_4_3,kujiale_0165,11.354283409514471,0,0.0,0.0
+kujiale_0165_50_1,kujiale_0165,3.1532456525302686,0,1.0,0.0
+kujiale_0165_29_1,kujiale_0165,4.394567118935355,0,0.0,0.0
+kujiale_0165_9_1,kujiale_0165,3.1585995567760885,0,0.0,0.0
+kujiale_0165_6_2,kujiale_0165,2.7471250892350163,1,1.0,0.6963204006574686
+kujiale_0165_46_3,kujiale_0165,6.11391219917519,0,0.0,0.0
+kujiale_0165_27_4,kujiale_0165,0.1116559245810509,1,1.0,0.9673276214215178
+kujiale_0128_fix_14_0,kujiale_0128,1.313993699711385,1,1.0,1.0
+kujiale_0128_fix_1_4,kujiale_0128,2.361511161296996,1,1.0,1.0
+kujiale_0128_fix_4_2,kujiale_0128,1.1043009140346842,1,1.0,1.0
+kujiale_0128_fix_13_1,kujiale_0128,5.117302223986329,0,0.0,0.0
+kujiale_0128_fix_1_3,kujiale_0128,3.895792848363158,0,0.0,0.0
+kujiale_0128_fix_11_4,kujiale_0128,2.6621055098399644,1,1.0,1.0
+kujiale_0128_fix_17_0,kujiale_0128,0.031216247619905248,1,1.0,1.0
+kujiale_0128_fix_14_3,kujiale_0128,0.5187483020381013,1,1.0,1.0
+kujiale_0128_fix_9_1,kujiale_0128,5.486179505869227,0,0.0,0.0
+kujiale_0128_fix_5_0,kujiale_0128,1.281956158217179,1,1.0,1.0
+kujiale_0128_fix_7_2,kujiale_0128,0.3541140310991837,1,1.0,1.0
+kujiale_0128_fix_19_2,kujiale_0128,6.1104926962290875,0,0.0,0.0
+kujiale_0128_fix_20_4,kujiale_0128,0.8048620841404709,1,1.0,1.0
+kujiale_0128_fix_0_0,kujiale_0128,5.771270709892611,0,0.0,0.0
+kujiale_0128_fix_12_1,kujiale_0128,6.703536502403986,0,0.0,0.0
+kujiale_0128_fix_6_3,kujiale_0128,0.9614181941729937,1,1.0,1.0
+kujiale_0128_fix_21_2,kujiale_0128,1.9757770866172724,1,1.0,1.0
+kujiale_0128_fix_3_2,kujiale_0128,4.6055136686034475,0,0.0,0.0
+kujiale_0128_fix_10_1,kujiale_0128,5.666697405915057,0,0.0,0.0
+kujiale_0128_fix_0_2,kujiale_0128,1.0063426577026997,1,1.0,1.0
+kujiale_0128_fix_1_1,kujiale_0128,2.068136083170321,1,1.0,1.0
+kujiale_0128_fix_3_3,kujiale_0128,5.785594973289717,0,0.0,0.0
+kujiale_0128_fix_12_0,kujiale_0128,4.211917205330669,0,0.0,0.0
+kujiale_0128_fix_11_3,kujiale_0128,2.269564293977474,1,1.0,1.0
+kujiale_0128_fix_1_0,kujiale_0128,4.749612085890035,0,1.0,0.0

eval_results/shards/shard2_serial/server.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c8d4853ac2c5b97eb4bd575fe54a94d837a1196f3b3661674bf5c0629e37adb
+size 74764

eval_results/shards/shard3_serial/eval.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92472e1e748de34f6fcc78130645597a78d89e200f3ee38f2c761dbded96b1a1
+size 7189047

eval_results/shards/shard3_serial/eval_cfg.py

@@ -0,0 +1,105 @@
+"""RyWorld eval cfg — VLNVerse coarse / val_unseen, H1 robot at 256x256.
+
+Forked from
+``third_party/vlnverse_emr/scripts/eval/configs/h1_cma_clip_cfg_vlnverse_coarse.py``
+with the minimum diffs needed for our model:
+
+  - ``agent.model_name='ryworld'`` so Agent.init dispatches to RyWorldAgent.
+  - ``split_data_types=['val_unseen']`` only (CMA cfg also includes val_seen).
+  - ``task_name`` carries our run identity for ``logs/<task>/`` separation.
+
+Everything else (camera 256x256, ``torso_link/h1_pano_camera_0``,
+``h1_vln_pointcloud.usd``, ``vln_pe`` env, ``flash`` mode, max_step=500,
+formal instruction) is kept identical to the CMA-CLIP baseline so our
+numbers go on the same table.
+
+Plugin registration: this cfg's top-level ``import
+ryworld.eval.vlnverse_plugin`` triggers ``@Agent.register('ryworld')`` as a
+side effect, which is what lets vlnverse_emr's dispatcher find our agent
+without a fork. ``importlib.util.spec_from_file_location`` (used by both
+``server.py`` and ``eval.py``) executes this module top-to-bottom, so by the
+time ``Agent.init`` runs the registration is already in place.
+"""
+
+from __future__ import annotations
+
+# ---------------------------------------------------------------------------
+# 1. Make the RyWorld package importable.
+#    vlnverse_emr's launchers run from inside ``third_party/vlnverse_emr``
+#    (their ``cwd``), so ``ryworld.*`` is not on ``sys.path`` by default. We
+#    walk up from this file: cfg -> configs -> vlnverse_plugin -> eval ->
+#    ryworld -> repo root.
+# ---------------------------------------------------------------------------
+import sys
+from pathlib import Path
+
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+# ---------------------------------------------------------------------------
+# 2. Import the plugin package — side effect: ``@Agent.register('ryworld')``.
+# ---------------------------------------------------------------------------
+import ryworld.eval.vlnverse_plugin  # noqa: E402, F401
+
+# ---------------------------------------------------------------------------
+# 3. Build the eval cfg using vlnverse_emr's pydantic config classes.
+# ---------------------------------------------------------------------------
+from vlnverse.configs.agent import AgentCfg  # type: ignore[import-not-found]  # noqa: E402
+from vlnverse.configs.evaluator import (  # type: ignore[import-not-found]  # noqa: E402
+    EnvCfg,
+    EvalCfg,
+    EvalDatasetCfg,
+    SceneCfg,
+    TaskCfg,
+)
+
+eval_cfg = EvalCfg(
+    agent=AgentCfg(
+        # Server port — keep clear of CMA(8087)/N1(8028) so we can run side-by-side.
+        server_port=8094,
+        model_name="ryworld",
+        # Dummy skeleton agent ignores ckpt_path; real wrapper will load this.
+        ckpt_path="/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+        model_settings={
+            "stop_threshold_override": 0.95,
+            "train_config_path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/stage1_discrete.yaml",
+            "train_config_overrides": ["/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/overlays/a100_4gpu_discrete.yaml"],
+        },
+    ),
+    env=EnvCfg(
+        env_type="vln_pe",
+        env_settings={
+            "use_fabric": False,
+            "headless": True,
+        },
+    ),
+    task=TaskCfg(
+        task_name="20260513T032631_step0015000_stop0.95_full835_thr095_shard3_serial",
+        task_settings={
+            "env_num": 1,
+            "use_distributed": False,
+            "proc_num": 1,
+            "max_step": 500,
+            "warm_up_step": 500,
+        },
+        scene=SceneCfg(
+            scene_type="kujiale",
+            scene_data_dir="data/scene_data/vlnverse",
+        ),
+        robot_name="h1",
+        robot_flash=True,
+        robot_usd_path="data/Embodiments/vln-pe/h1/h1_vln_pointcloud.usd",
+        camera_resolution=[256, 256],
+        camera_prim_path="torso_link/h1_pano_camera_0",
+    ),
+    dataset=EvalDatasetCfg(
+        dataset_type="kujiale",
+        dataset_settings={
+            "base_data_dir": "data/vlnverse/raw_data/vlnverse/coarse_shard3",
+            "split_data_types": ["val_unseen"],
+            "filter_stairs": False,
+            "instruction_type": "formal",
+        },
+    ),
+)

eval_results/shards/shard3_serial/eval_result.log

@@ -0,0 +1,16 @@
+[split:val_unseen] total 104 data
+############[val_unseen]#############
+TL = 717.9381120914196 / 104 = 6.9033
+NE = 446.72012255505564 / 104 = 4.2954
+FR = 0 / 104 = 0.0%
+StR = 0 / 104 = 0%
+OS = 57.0 / 104 = 54.81%
+SR = 49.0 / 104 = 47.12%
+SPL = 47.05632141805367 / 104 = 45.25%
+detail:
+[reach_goal]:49
+[success]:49
+[not_reach_goal]:53
+[exceed_total_max_step]:2
+[fall]:0
+##########################

eval_results/shards/shard3_serial/meta.json

@@ -0,0 +1,161 @@
+{
+  "eval_id": "20260513T032631_step0015000_stop0.95_full835_thr095_shard3_serial",
+  "tag": "full835_thr095_shard3_serial",
+  "timestamp_recorded": "2026-05-13T17:11:20+0800",
+  "ckpt": {
+    "path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+    "step": 15000,
+    "size_bytes": 2812658821
+  },
+  "code_state": {
+    "git_hash": "a4fa65370358d6f251e158f03aced7d30f167da6",
+    "branch": "feat/batched-trunk-forward",
+    "dirty": true,
+    "dirty_files": [
+      "docs/eval_ledger.jsonl",
+      "run_train.sh",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD0_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD1_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD2_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD3_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD4_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD5_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD6_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD7_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu0_thr091_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu1_thr092_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu2_thr093_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu3_thr094_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu4_thr095_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu5_thr096_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu6_thr097_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu7_thr098_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v7_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v8_port8092.py",
+      "scripts/eval/run_eval_structured.sh",
+      "third_party/vlnverse_emr",
+      "yworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen.py"
+    ],
+    "dirty_patch_md5": "aa7101e2b529b2f5596267cd1e8073ee"
+  },
+  "eval_config": {
+    "framework": "vlnverse_emr",
+    "base_cfg": "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD3_port8094.py",
+    "derived_cfg": "checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/evals/20260513T032631_step0015000_stop0.95_full835_thr095_shard3_serial/eval_cfg.py",
+    "stop_threshold": 0.95,
+    "task_name": "20260513T032631_step0015000_stop0.95_full835_thr095_shard3_serial"
+  },
+  "data_state": {
+    "scenes_used": [
+      "kujiale_0102",
+      "kujiale_0114",
+      "kujiale_0207",
+      "kujiale_0208",
+      "kujiale_0226"
+    ],
+    "scenes_with_hdr": [
+      "kujiale_0102",
+      "kujiale_0114",
+      "kujiale_0207",
+      "kujiale_0208",
+      "kujiale_0226"
+    ],
+    "scenes_missing_hdr": []
+  },
+  "hardware": {
+    "host": "di-20260502152021-n5hs7",
+    "gpus": [
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      }
+    ],
+    "count": 8
+  },
+  "results_aggregate": {
+    "TL": 6.9033,
+    "NE": 4.2954,
+    "FR": 0.0,
+    "StR": 0.0,
+    "OS": 0.5481,
+    "SR": 0.47119999999999995,
+    "SPL": 0.4525,
+    "n_reach_goal": 49,
+    "n_not_reach_goal": 53,
+    "n_success": 49,
+    "n_fall": 0,
+    "n_total": 104
+  },
+  "results_per_scene": {
+    "kujiale_0226": {
+      "n": 25,
+      "SR": 0.56,
+      "avg_NE": 4.338929470504046,
+      "avg_SPL": 0.5292428458521876
+    },
+    "kujiale_0208": {
+      "n": 25,
+      "SR": 0.36,
+      "avg_NE": 4.535285835438256,
+      "avg_SPL": 0.36
+    },
+    "kujiale_0207": {
+      "n": 23,
+      "SR": 0.34782608695652173,
+      "avg_NE": 5.630842382057829,
+      "avg_SPL": 0.33256955591271564
+    },
+    "kujiale_0114": {
+      "n": 7,
+      "SR": 0.14285714285714285,
+      "avg_NE": 6.128924782437337,
+      "avg_SPL": 0.14285714285714285
+    },
+    "kujiale_0102": {
+      "n": 23,
+      "SR": 0.7391304347826086,
+      "avg_NE": 1.9255885534016466,
+      "avg_SPL": 0.7033108906850656
+    }
+  },
+  "stop_head_stats": {
+    "n_samples": 22136,
+    "stop_prob_min": 0.279,
+    "stop_prob_p50": 0.733,
+    "stop_prob_p90": 0.896,
+    "stop_prob_max": 0.977,
+    "stop_prob_mean": 0.7305024846404101,
+    "fire_rate_pathA": 0.022090711962414168,
+    "fire_rate_pathB": 0.006098662811709433,
+    "no_stop_rate": 0.9718106252258764
+  },
+  "n_episodes_recorded": 103
+}

eval_results/shards/shard3_serial/per_episode.csv

@@ -0,0 +1,104 @@
+episode_id,scene,NE,success,osr,SPL
+kujiale_0226_12_1,kujiale_0226,2.699860735892533,1,1.0,1.0
+kujiale_0226_49_3,kujiale_0226,9.026704140957465,0,0.0,0.0
+kujiale_0226_20_4,kujiale_0226,1.797690719911934,1,1.0,1.0
+kujiale_0226_41_0,kujiale_0226,1.9789351098482757,1,1.0,1.0
+kujiale_0226_19_1,kujiale_0226,2.551102199675326,1,1.0,1.0
+kujiale_0226_83_3,kujiale_0226,0.9297664984509062,1,1.0,1.0
+kujiale_0226_78_2,kujiale_0226,2.688230365283556,1,1.0,0.8709315568036395
+kujiale_0226_17_3,kujiale_0226,7.76272632335025,0,0.0,0.0
+kujiale_0226_98_0,kujiale_0226,1.830652091238978,1,1.0,1.0
+kujiale_0226_65_2,kujiale_0226,4.281912335942562,0,0.0,0.0
+kujiale_0226_60_0,kujiale_0226,1.118843530521249,1,1.0,0.8623832867387594
+kujiale_0226_10_0,kujiale_0226,4.859473017033209,0,0.0,0.0
+kujiale_0226_55_1,kujiale_0226,9.844562792047865,0,0.0,0.0
+kujiale_0226_101_0,kujiale_0226,2.7992734424807657,1,1.0,0.8528491198722784
+kujiale_0226_7_3,kujiale_0226,4.285048314136865,0,0.0,0.0
+kujiale_0226_61_0,kujiale_0226,1.2834080300557673,1,1.0,1.0
+kujiale_0226_97_1,kujiale_0226,7.6694025732562245,0,0.0,0.0
+kujiale_0226_42_2,kujiale_0226,4.038140766820334,0,1.0,0.0
+kujiale_0226_105_3,kujiale_0226,1.7657338763413346,1,1.0,0.7809186815292362
+kujiale_0226_67_4,kujiale_0226,8.98112959739926,0,0.0,0.0
+kujiale_0226_4_4,kujiale_0226,9.690141004338923,0,0.0,0.0
+kujiale_0226_37_4,kujiale_0226,2.933453892077824,1,1.0,1.0
+kujiale_0226_79_3,kujiale_0226,11.387483002671294,0,0.0,0.0
+kujiale_0226_46_3,kujiale_0226,0.48543573440815707,1,1.0,0.863988501360777
+kujiale_0226_57_1,kujiale_0226,1.7841266684602939,1,1.0,1.0
+kujiale_0208_15_2,kujiale_0208,4.8233278492193605,0,0.0,0.0
+kujiale_0208_34_0,kujiale_0208,4.188540080158509,0,0.0,0.0
+kujiale_0208_35_4,kujiale_0208,5.079366798035417,0,0.0,0.0
+kujiale_0208_14_1,kujiale_0208,6.695019285572422,0,0.0,0.0
+kujiale_0208_41_0,kujiale_0208,8.469049472862856,0,0.0,0.0
+kujiale_0208_23_4,kujiale_0208,6.07525610481124,0,0.0,0.0
+kujiale_0208_28_2,kujiale_0208,10.038437387409848,0,0.0,0.0
+kujiale_0208_6_2,kujiale_0208,2.4317303425792725,1,1.0,1.0
+kujiale_0208_44_1,kujiale_0208,0.5655557352821946,1,1.0,1.0
+kujiale_0208_42_1,kujiale_0208,6.681077127970111,0,0.0,0.0
+kujiale_0208_24_0,kujiale_0208,6.664456438283543,0,0.0,0.0
+kujiale_0208_1_3,kujiale_0208,2.687947118518917,1,1.0,1.0
+kujiale_0208_30_0,kujiale_0208,2.2526075520949855,1,1.0,1.0
+kujiale_0208_26_0,kujiale_0208,5.666754386937908,0,0.0,0.0
+kujiale_0208_49_1,kujiale_0208,1.1493853001979775,1,1.0,1.0
+kujiale_0208_43_3,kujiale_0208,3.35363003472937,0,0.0,0.0
+kujiale_0208_11_2,kujiale_0208,4.292876933233738,0,0.0,0.0
+kujiale_0208_38_4,kujiale_0208,9.64553192901746,0,0.0,0.0
+kujiale_0208_20_1,kujiale_0208,3.6051555307003693,0,1.0,0.0
+kujiale_0208_10_0,kujiale_0208,1.9748993449172871,1,1.0,1.0
+kujiale_0208_33_1,kujiale_0208,0.48659892473051886,1,1.0,1.0
+kujiale_0208_0_0,kujiale_0208,0.8246881075099005,1,1.0,1.0
+kujiale_0208_12_3,kujiale_0208,9.109046690543957,0,0.0,0.0
+kujiale_0208_16_1,kujiale_0208,2.674607538299321,1,1.0,1.0
+kujiale_0208_17_2,kujiale_0208,3.946599872339903,0,0.0,0.0
+kujiale_0207_19_3,kujiale_0207,10.003210513352682,0,0.0,0.0
+kujiale_0207_29_2,kujiale_0207,9.69090130203633,0,0.0,0.0
+kujiale_0207_77_1,kujiale_0207,5.069625078379125,0,0.0,0.0
+kujiale_0207_48_1,kujiale_0207,14.118456306607646,0,0.0,0.0
+kujiale_0207_56_1,kujiale_0207,0.5668540792117858,1,1.0,0.9553315413852012
+kujiale_0207_10_2,kujiale_0207,10.347844668136531,0,0.0,0.0
+kujiale_0207_67_2,kujiale_0207,2.063386332964216,1,1.0,1.0
+kujiale_0207_5_3,kujiale_0207,6.519225940634539,0,1.0,0.0
+kujiale_0207_21_0,kujiale_0207,0.8358835920943863,1,1.0,1.0
+kujiale_0207_53_0,kujiale_0207,6.512274834611155,0,0.0,0.0
+kujiale_0207_33_1,kujiale_0207,12.329957441662778,0,0.0,0.0
+kujiale_0207_66_0,kujiale_0207,5.817983591770667,0,0.0,0.0
+kujiale_0207_73_1,kujiale_0207,6.2316581028104405,0,0.0,0.0
+kujiale_0207_61_0,kujiale_0207,8.586153323106052,0,0.0,0.0
+kujiale_0207_57_3,kujiale_0207,6.239034595148907,0,0.0,0.0
+kujiale_0207_17_4,kujiale_0207,0.3337308381782926,1,1.0,0.9830564668052867
+kujiale_0207_64_2,kujiale_0207,0.24843205169850216,1,1.0,0.7107117778019709
+kujiale_0207_58_1,kujiale_0207,1.3543494940833682,1,1.0,1.0
+kujiale_0207_80_4,kujiale_0207,8.75396776531336,0,1.0,0.0
+kujiale_0207_0_4,kujiale_0207,5.058640422243232,0,0.0,0.0
+kujiale_0207_39_2,kujiale_0207,4.645423131533198,0,1.0,0.0
+kujiale_0207_47_3,kujiale_0207,1.2123913288896215,1,1.0,1.0
+kujiale_0207_52_0,kujiale_0207,2.9699900528632734,1,1.0,1.0
+kujiale_0114_98_0,kujiale_0114,3.3376698247010017,0,0.0,0.0
+kujiale_0114_111_4,kujiale_0114,4.774217840667842,0,0.0,0.0
+kujiale_0114_94_2,kujiale_0114,1.693736759057279,1,1.0,1.0
+kujiale_0114_77_2,kujiale_0114,5.816149213670121,0,0.0,0.0
+kujiale_0114_74_4,kujiale_0114,11.318388192231255,0,0.0,0.0
+kujiale_0114_123_4,kujiale_0114,10.441168618316155,0,0.0,0.0
+kujiale_0114_134_2,kujiale_0114,5.521143028417705,0,0.0,0.0
+kujiale_0102_51_3,kujiale_0102,4.240156209633976,0,1.0,0.0
+kujiale_0102_5_3,kujiale_0102,0.4596704371583703,1,1.0,1.0
+kujiale_0102_30_0,kujiale_0102,0.22049633443955333,1,1.0,1.0
+kujiale_0102_37_1,kujiale_0102,1.4463988846738434,1,1.0,0.8251787693887231
+kujiale_0102_33_2,kujiale_0102,0.8087253840881719,1,1.0,0.6846847134554074
+kujiale_0102_48_4,kujiale_0102,2.403489488780312,1,1.0,1.0
+kujiale_0102_1_0,kujiale_0102,4.574795703104781,0,0.0,0.0
+kujiale_0102_10_0,kujiale_0102,3.488794593076284,0,0.0,0.0
+kujiale_0102_31_4,kujiale_0102,1.2578293461790993,1,1.0,1.0
+kujiale_0102_6_3,kujiale_0102,2.715623291754815,1,1.0,1.0
+kujiale_0102_17_3,kujiale_0102,0.6183007637359813,1,1.0,1.0
+kujiale_0102_40_4,kujiale_0102,5.945505636486258,0,1.0,0.0
+kujiale_0102_36_1,kujiale_0102,0.34150117526462537,1,1.0,1.0
+kujiale_0102_53_0,kujiale_0102,0.4600993688820919,1,1.0,1.0
+kujiale_0102_4_3,kujiale_0102,0.9558154076590545,1,1.0,0.7722246744125937
+kujiale_0102_56_2,kujiale_0102,0.4146459968632296,1,1.0,1.0
+kujiale_0102_27_0,kujiale_0102,0.4113550137833243,1,1.0,1.0
+kujiale_0102_26_1,kujiale_0102,4.723994237504759,0,0.0,0.0
+kujiale_0102_54_1,kujiale_0102,3.633867200364762,0,1.0,0.0
+kujiale_0102_7_4,kujiale_0102,1.8377293294682557,1,1.0,1.0
+kujiale_0102_11_2,kujiale_0102,0.9725366186728053,1,1.0,0.8940623284997845
+kujiale_0102_38_2,kujiale_0102,1.3363462915941082,1,1.0,1.0
+kujiale_0102_23_4,kujiale_0102,1.0208600150694023,1,1.0,1.0

eval_results/shards/shard3_serial/server.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5ddc1947e08db48caf50939e5165e411df524960c6cb9b4d1ed9305dd9d55bc
+size 99786

eval_results/shards/shard4_serial/eval.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a39559a7acaa8f5e5b974e39ba9686f98a869d3275eea73c0d3443cd0eb4b381
+size 6060440

eval_results/shards/shard4_serial/eval_cfg.py

@@ -0,0 +1,105 @@
+"""RyWorld eval cfg — VLNVerse coarse / val_unseen, H1 robot at 256x256.
+
+Forked from
+``third_party/vlnverse_emr/scripts/eval/configs/h1_cma_clip_cfg_vlnverse_coarse.py``
+with the minimum diffs needed for our model:
+
+  - ``agent.model_name='ryworld'`` so Agent.init dispatches to RyWorldAgent.
+  - ``split_data_types=['val_unseen']`` only (CMA cfg also includes val_seen).
+  - ``task_name`` carries our run identity for ``logs/<task>/`` separation.
+
+Everything else (camera 256x256, ``torso_link/h1_pano_camera_0``,
+``h1_vln_pointcloud.usd``, ``vln_pe`` env, ``flash`` mode, max_step=500,
+formal instruction) is kept identical to the CMA-CLIP baseline so our
+numbers go on the same table.
+
+Plugin registration: this cfg's top-level ``import
+ryworld.eval.vlnverse_plugin`` triggers ``@Agent.register('ryworld')`` as a
+side effect, which is what lets vlnverse_emr's dispatcher find our agent
+without a fork. ``importlib.util.spec_from_file_location`` (used by both
+``server.py`` and ``eval.py``) executes this module top-to-bottom, so by the
+time ``Agent.init`` runs the registration is already in place.
+"""
+
+from __future__ import annotations
+
+# ---------------------------------------------------------------------------
+# 1. Make the RyWorld package importable.
+#    vlnverse_emr's launchers run from inside ``third_party/vlnverse_emr``
+#    (their ``cwd``), so ``ryworld.*`` is not on ``sys.path`` by default. We
+#    walk up from this file: cfg -> configs -> vlnverse_plugin -> eval ->
+#    ryworld -> repo root.
+# ---------------------------------------------------------------------------
+import sys
+from pathlib import Path
+
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+# ---------------------------------------------------------------------------
+# 2. Import the plugin package — side effect: ``@Agent.register('ryworld')``.
+# ---------------------------------------------------------------------------
+import ryworld.eval.vlnverse_plugin  # noqa: E402, F401
+
+# ---------------------------------------------------------------------------
+# 3. Build the eval cfg using vlnverse_emr's pydantic config classes.
+# ---------------------------------------------------------------------------
+from vlnverse.configs.agent import AgentCfg  # type: ignore[import-not-found]  # noqa: E402
+from vlnverse.configs.evaluator import (  # type: ignore[import-not-found]  # noqa: E402
+    EnvCfg,
+    EvalCfg,
+    EvalDatasetCfg,
+    SceneCfg,
+    TaskCfg,
+)
+
+eval_cfg = EvalCfg(
+    agent=AgentCfg(
+        # Server port — keep clear of CMA(8087)/N1(8028) so we can run side-by-side.
+        server_port=8095,
+        model_name="ryworld",
+        # Dummy skeleton agent ignores ckpt_path; real wrapper will load this.
+        ckpt_path="/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+        model_settings={
+            "stop_threshold_override": 0.95,
+            "train_config_path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/stage1_discrete.yaml",
+            "train_config_overrides": ["/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/overlays/a100_4gpu_discrete.yaml"],
+        },
+    ),
+    env=EnvCfg(
+        env_type="vln_pe",
+        env_settings={
+            "use_fabric": False,
+            "headless": True,
+        },
+    ),
+    task=TaskCfg(
+        task_name="20260513T032731_step0015000_stop0.95_full835_thr095_shard4_serial",
+        task_settings={
+            "env_num": 1,
+            "use_distributed": False,
+            "proc_num": 1,
+            "max_step": 500,
+            "warm_up_step": 500,
+        },
+        scene=SceneCfg(
+            scene_type="kujiale",
+            scene_data_dir="data/scene_data/vlnverse",
+        ),
+        robot_name="h1",
+        robot_flash=True,
+        robot_usd_path="data/Embodiments/vln-pe/h1/h1_vln_pointcloud.usd",
+        camera_resolution=[256, 256],
+        camera_prim_path="torso_link/h1_pano_camera_0",
+    ),
+    dataset=EvalDatasetCfg(
+        dataset_type="kujiale",
+        dataset_settings={
+            "base_data_dir": "data/vlnverse/raw_data/vlnverse/coarse_shard4",
+            "split_data_types": ["val_unseen"],
+            "filter_stairs": False,
+            "instruction_type": "formal",
+        },
+    ),
+)

eval_results/shards/shard4_serial/eval_result.log

@@ -0,0 +1,16 @@
+[split:val_unseen] total 104 data
+############[val_unseen]#############
+TL = 556.99267312641 / 104 = 5.3557
+NE = 332.4797110170293 / 104 = 3.1969
+FR = 0 / 104 = 0.0%
+StR = 0 / 104 = 0%
+OS = 70.0 / 104 = 67.31%
+SR = 59.0 / 104 = 56.730000000000004%
+SPL = 57.10930306111026 / 104 = 54.910000000000004%
+detail:
+[reach_goal]:59
+[success]:58
+[not_reach_goal]:43
+[exceed_total_max_step]:3
+[fall]:0
+##########################

eval_results/shards/shard4_serial/meta.json

@@ -0,0 +1,161 @@
+{
+  "eval_id": "20260513T032731_step0015000_stop0.95_full835_thr095_shard4_serial",
+  "tag": "full835_thr095_shard4_serial",
+  "timestamp_recorded": "2026-05-13T16:19:07+0800",
+  "ckpt": {
+    "path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+    "step": 15000,
+    "size_bytes": 2812658821
+  },
+  "code_state": {
+    "git_hash": "a4fa65370358d6f251e158f03aced7d30f167da6",
+    "branch": "feat/batched-trunk-forward",
+    "dirty": true,
+    "dirty_files": [
+      "docs/eval_ledger.jsonl",
+      "run_train.sh",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD0_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD1_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD2_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD3_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD4_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD5_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD6_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD7_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu0_thr091_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu1_thr092_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu2_thr093_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu3_thr094_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu4_thr095_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu5_thr096_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu6_thr097_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu7_thr098_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v7_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v8_port8092.py",
+      "scripts/eval/run_eval_structured.sh",
+      "third_party/vlnverse_emr",
+      "yworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen.py"
+    ],
+    "dirty_patch_md5": "aa7101e2b529b2f5596267cd1e8073ee"
+  },
+  "eval_config": {
+    "framework": "vlnverse_emr",
+    "base_cfg": "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD4_port8095.py",
+    "derived_cfg": "checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/evals/20260513T032731_step0015000_stop0.95_full835_thr095_shard4_serial/eval_cfg.py",
+    "stop_threshold": 0.95,
+    "task_name": "20260513T032731_step0015000_stop0.95_full835_thr095_shard4_serial"
+  },
+  "data_state": {
+    "scenes_used": [
+      "kujiale_0030",
+      "kujiale_0031",
+      "kujiale_0036",
+      "kujiale_0066",
+      "kujiale_0114"
+    ],
+    "scenes_with_hdr": [
+      "kujiale_0030",
+      "kujiale_0031",
+      "kujiale_0036",
+      "kujiale_0066",
+      "kujiale_0114"
+    ],
+    "scenes_missing_hdr": []
+  },
+  "hardware": {
+    "host": "di-20260502152021-n5hs7",
+    "gpus": [
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      }
+    ],
+    "count": 8
+  },
+  "results_aggregate": {
+    "TL": 5.3557,
+    "NE": 3.1969,
+    "FR": 0.0,
+    "StR": 0.0,
+    "OS": 0.6731,
+    "SR": 0.5673,
+    "SPL": 0.5491,
+    "n_reach_goal": 59,
+    "n_not_reach_goal": 43,
+    "n_success": 58,
+    "n_fall": 0,
+    "n_total": 104
+  },
+  "results_per_scene": {
+    "kujiale_0114": {
+      "n": 12,
+      "SR": 0.4166666666666667,
+      "avg_NE": 3.9252291526806573,
+      "avg_SPL": 0.4166666666666667
+    },
+    "kujiale_0066": {
+      "n": 17,
+      "SR": 0.6470588235294118,
+      "avg_NE": 2.9718380968099227,
+      "avg_SPL": 0.6085980775492333
+    },
+    "kujiale_0036": {
+      "n": 25,
+      "SR": 0.52,
+      "avg_NE": 3.5283700892273417,
+      "avg_SPL": 0.4943872645627582
+    },
+    "kujiale_0031": {
+      "n": 25,
+      "SR": 0.56,
+      "avg_NE": 3.153615188659501,
+      "avg_SPL": 0.56
+    },
+    "kujiale_0030": {
+      "n": 24,
+      "SR": 0.625,
+      "avg_NE": 2.790199554025632,
+      "avg_SPL": 0.6001439220293476
+    }
+  },
+  "stop_head_stats": {
+    "n_samples": 19236,
+    "stop_prob_min": 0.274,
+    "stop_prob_p50": 0.711,
+    "stop_prob_p90": 0.882,
+    "stop_prob_max": 0.979,
+    "stop_prob_mean": 0.7210352464129736,
+    "fire_rate_pathA": 0.027552505718444583,
+    "fire_rate_pathB": 0.006654190060303597,
+    "no_stop_rate": 0.9657933042212519
+  },
+  "n_episodes_recorded": 103
+}

eval_results/shards/shard4_serial/per_episode.csv

@@ -0,0 +1,104 @@
+episode_id,scene,NE,success,osr,SPL
+kujiale_0114_20_3,kujiale_0114,8.4491552703076,0,0.0,0.0
+kujiale_0114_93_2,kujiale_0114,0.5715126196899469,1,1.0,1.0
+kujiale_0114_133_0,kujiale_0114,4.930240763348038,0,0.0,0.0
+kujiale_0114_131_1,kujiale_0114,0.38957559891708904,1,1.0,1.0
+kujiale_0114_82_3,kujiale_0114,10.224838863393337,0,0.0,0.0
+kujiale_0114_88_2,kujiale_0114,5.339168947696867,0,0.0,0.0
+kujiale_0114_67_1,kujiale_0114,2.248189772352827,1,1.0,1.0
+kujiale_0114_11_2,kujiale_0114,3.465643740290391,0,1.0,0.0
+kujiale_0114_138_3,kujiale_0114,4.174692810473763,0,0.0,0.0
+kujiale_0114_17_4,kujiale_0114,3.2055532508131344,0,0.0,0.0
+kujiale_0114_107_1,kujiale_0114,1.2811206857534694,1,1.0,1.0
+kujiale_0114_78_4,kujiale_0114,2.8230575091314245,1,1.0,1.0
+kujiale_0066_15_2,kujiale_0066,2.5064636446595223,1,1.0,1.0
+kujiale_0066_18_4,kujiale_0066,2.704230492155911,1,1.0,0.6388335312866373
+kujiale_0066_24_0,kujiale_0066,3.5061054794607003,0,0.0,0.0
+kujiale_0066_27_3,kujiale_0066,2.970719360555563,1,1.0,0.8689187039711138
+kujiale_0066_22_2,kujiale_0066,0.9275476211111354,1,1.0,1.0
+kujiale_0066_35_4,kujiale_0066,3.209687590840927,0,0.0,0.0
+kujiale_0066_43_3,kujiale_0066,1.5947263230870334,1,1.0,1.0
+kujiale_0066_6_2,kujiale_0066,2.14324064260706,1,1.0,1.0
+kujiale_0066_40_0,kujiale_0066,2.9554537034796104,1,1.0,1.0
+kujiale_0066_17_2,kujiale_0066,6.269673280965588,0,0.0,0.0
+kujiale_0066_4_1,kujiale_0066,4.012393136546961,0,1.0,0.0
+kujiale_0066_14_1,kujiale_0066,3.616718505393394,0,0.0,0.0
+kujiale_0066_44_1,kujiale_0066,5.247048650649337,0,0.0,0.0
+kujiale_0066_2_3,kujiale_0066,2.6859787190741664,1,1.0,0.8945703715965809
+kujiale_0066_16_1,kujiale_0066,1.6147696497007786,1,1.0,0.9438447114826353
+kujiale_0066_28_2,kujiale_0066,2.5920059168836196,1,1.0,1.0
+kujiale_0066_33_1,kujiale_0066,1.964484928597389,1,1.0,1.0
+kujiale_0036_2_3,kujiale_0036,2.6012731426822104,1,1.0,0.7667088041952344
+kujiale_0036_38_3,kujiale_0036,1.963107138235835,1,1.0,0.9720594177530585
+kujiale_0036_35_2,kujiale_0036,0.8359444555016114,1,1.0,1.0
+kujiale_0036_45_4,kujiale_0036,8.208507463775979,0,0.0,0.0
+kujiale_0036_15_3,kujiale_0036,3.999954556718976,0,1.0,0.0
+kujiale_0036_23_1,kujiale_0036,6.043885681080589,0,0.0,0.0
+kujiale_0036_49_0,kujiale_0036,1.6799414857902955,1,1.0,1.0
+kujiale_0036_53_0,kujiale_0036,5.138120045202638,0,0.0,0.0
+kujiale_0036_22_2,kujiale_0036,0.5470583394591437,1,1.0,0.8923216853459757
+kujiale_0036_9_1,kujiale_0036,5.918162500932902,0,1.0,0.0
+kujiale_0036_46_3,kujiale_0036,1.7495830449046055,1,1.0,1.0
+kujiale_0036_31_1,kujiale_0036,3.9892892901897485,0,1.0,0.0
+kujiale_0036_24_0,kujiale_0036,0.297645647129644,1,1.0,1.0
+kujiale_0036_47_2,kujiale_0036,1.2894515569395404,1,1.0,1.0
+kujiale_0036_54_0,kujiale_0036,2.917426345446369,1,1.0,1.0
+kujiale_0036_5_2,kujiale_0036,9.319695981246184,0,0.0,0.0
+kujiale_0036_40_0,kujiale_0036,6.1333532430052315,0,0.0,0.0
+kujiale_0036_42_4,kujiale_0036,1.109546915023238,1,1.0,1.0
+kujiale_0036_28_3,kujiale_0036,0.6033904888470073,1,1.0,1.0
+kujiale_0036_25_0,kujiale_0036,5.145782147567935,0,0.0,0.0
+kujiale_0036_1_0,kujiale_0036,5.83777726820394,0,0.0,0.0
+kujiale_0036_29_4,kujiale_0036,4.605211837059694,0,0.0,0.0
+kujiale_0036_20_1,kujiale_0036,1.6883968430406329,1,1.0,1.0
+kujiale_0036_55_0,kujiale_0036,2.791724224958323,1,1.0,0.7285917067746887
+kujiale_0036_12_2,kujiale_0036,3.7950225877412826,0,0.0,0.0
+kujiale_0031_13_3,kujiale_0031,1.094703935063376,1,1.0,1.0
+kujiale_0031_14_0,kujiale_0031,3.721661905618094,0,0.0,0.0
+kujiale_0031_5_4,kujiale_0031,2.049454197241087,1,1.0,1.0
+kujiale_0031_9_0,kujiale_0031,1.537699107051595,1,1.0,1.0
+kujiale_0031_20_4,kujiale_0031,3.8839131244397023,0,0.0,0.0
+kujiale_0031_3_1,kujiale_0031,5.740003951080485,0,0.0,0.0
+kujiale_0031_2_4,kujiale_0031,2.5290311083359316,1,1.0,1.0
+kujiale_0031_19_3,kujiale_0031,5.075537823990019,0,0.0,0.0
+kujiale_0031_2_2,kujiale_0031,3.371777024212601,0,0.0,0.0
+kujiale_0031_1_3,kujiale_0031,2.7563637547664674,1,1.0,1.0
+kujiale_0031_25_0,kujiale_0031,4.797116596988436,0,0.0,0.0
+kujiale_0031_18_2,kujiale_0031,6.533442183155678,0,0.0,0.0
+kujiale_0031_5_0,kujiale_0031,3.0182771552124805,0,1.0,0.0
+kujiale_0031_20_1,kujiale_0031,2.6105005776167087,1,1.0,1.0
+kujiale_0031_4_1,kujiale_0031,2.6497392832948754,1,1.0,1.0
+kujiale_0031_22_4,kujiale_0031,2.703563041426141,1,1.0,1.0
+kujiale_0031_22_0,kujiale_0031,5.462057736483102,0,0.0,0.0
+kujiale_0031_11_4,kujiale_0031,2.389458497348246,1,1.0,1.0
+kujiale_0031_7_1,kujiale_0031,2.65743106664255,1,1.0,1.0
+kujiale_0031_14_3,kujiale_0031,4.0279844511903296,0,0.0,0.0
+kujiale_0031_15_3,kujiale_0031,1.114849350350237,1,1.0,1.0
+kujiale_0031_17_3,kujiale_0031,2.3969447541160678,1,1.0,1.0
+kujiale_0031_23_3,kujiale_0031,2.377210388112735,1,1.0,1.0
+kujiale_0031_5_1,kujiale_0031,3.731717899103073,0,0.0,0.0
+kujiale_0031_2_3,kujiale_0031,0.6099408036475038,1,1.0,1.0
+kujiale_0030_25_0,kujiale_0030,0.7346146406408202,1,1.0,1.0
+kujiale_0030_15_3,kujiale_0030,1.0061885713382077,1,1.0,0.9278288787481853
+kujiale_0030_3_4,kujiale_0030,2.8224672643579467,1,1.0,0.6257834472843516
+kujiale_0030_7_1,kujiale_0030,2.503801571484958,1,1.0,1.0
+kujiale_0030_26_0,kujiale_0030,3.115324251402958,0,1.0,0.0
+kujiale_0030_7_4,kujiale_0030,5.124380774537448,0,1.0,0.0
+kujiale_0030_12_2,kujiale_0030,3.1439921466614003,0,1.0,0.0
+kujiale_0030_10_2,kujiale_0030,7.259337574976993,0,1.0,0.0
+kujiale_0030_0_1,kujiale_0030,3.544674003274622,0,1.0,0.0
+kujiale_0030_13_3,kujiale_0030,2.28627793732491,1,1.0,1.0
+kujiale_0030_11_4,kujiale_0030,1.446113155262322,1,1.0,1.0
+kujiale_0030_19_3,kujiale_0030,2.41418440303642,1,1.0,1.0
+kujiale_0030_9_0,kujiale_0030,0.636443518915264,1,1.0,1.0
+kujiale_0030_4_1,kujiale_0030,3.5255903243732405,0,0.0,0.0
+kujiale_0030_9_2,kujiale_0030,6.713593588500706,0,0.0,0.0
+kujiale_0030_20_1,kujiale_0030,0.6448320694977903,1,1.0,1.0
+kujiale_0030_5_4,kujiale_0030,6.0847774968253505,0,0.0,0.0
+kujiale_0030_2_2,kujiale_0030,2.4290674049344196,1,1.0,1.0
+kujiale_0030_11_3,kujiale_0030,4.0385391310521435,0,0.0,0.0
+kujiale_0030_17_4,kujiale_0030,1.758451037457763,1,1.0,1.0
+kujiale_0030_22_4,kujiale_0030,2.636245584013792,1,1.0,0.8498418026718056
+kujiale_0030_5_0,kujiale_0030,1.5945001791046804,1,1.0,1.0
+kujiale_0030_12_1,kujiale_0030,0.9825265053313049,1,1.0,1.0
+kujiale_0030_24_0,kujiale_0030,0.5188661623096896,1,1.0,1.0

eval_results/shards/shard4_serial/server.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3f3b80520146099d77febd02f405892268b21b506186d81b43ec1e98a62626b
+size 84233

eval_results/shards/shard5/eval.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:054bee0150710fce2bc043bb5ccf87cfc9c5aa84f7c4bc77f07b00d0f9c254ab
+size 5006227

eval_results/shards/shard5/eval_cfg.py

@@ -0,0 +1,105 @@
+"""RyWorld eval cfg — VLNVerse coarse / val_unseen, H1 robot at 256x256.
+
+Forked from
+``third_party/vlnverse_emr/scripts/eval/configs/h1_cma_clip_cfg_vlnverse_coarse.py``
+with the minimum diffs needed for our model:
+
+  - ``agent.model_name='ryworld'`` so Agent.init dispatches to RyWorldAgent.
+  - ``split_data_types=['val_unseen']`` only (CMA cfg also includes val_seen).
+  - ``task_name`` carries our run identity for ``logs/<task>/`` separation.
+
+Everything else (camera 256x256, ``torso_link/h1_pano_camera_0``,
+``h1_vln_pointcloud.usd``, ``vln_pe`` env, ``flash`` mode, max_step=500,
+formal instruction) is kept identical to the CMA-CLIP baseline so our
+numbers go on the same table.
+
+Plugin registration: this cfg's top-level ``import
+ryworld.eval.vlnverse_plugin`` triggers ``@Agent.register('ryworld')`` as a
+side effect, which is what lets vlnverse_emr's dispatcher find our agent
+without a fork. ``importlib.util.spec_from_file_location`` (used by both
+``server.py`` and ``eval.py``) executes this module top-to-bottom, so by the
+time ``Agent.init`` runs the registration is already in place.
+"""
+
+from __future__ import annotations
+
+# ---------------------------------------------------------------------------
+# 1. Make the RyWorld package importable.
+#    vlnverse_emr's launchers run from inside ``third_party/vlnverse_emr``
+#    (their ``cwd``), so ``ryworld.*`` is not on ``sys.path`` by default. We
+#    walk up from this file: cfg -> configs -> vlnverse_plugin -> eval ->
+#    ryworld -> repo root.
+# ---------------------------------------------------------------------------
+import sys
+from pathlib import Path
+
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+# ---------------------------------------------------------------------------
+# 2. Import the plugin package — side effect: ``@Agent.register('ryworld')``.
+# ---------------------------------------------------------------------------
+import ryworld.eval.vlnverse_plugin  # noqa: E402, F401
+
+# ---------------------------------------------------------------------------
+# 3. Build the eval cfg using vlnverse_emr's pydantic config classes.
+# ---------------------------------------------------------------------------
+from vlnverse.configs.agent import AgentCfg  # type: ignore[import-not-found]  # noqa: E402
+from vlnverse.configs.evaluator import (  # type: ignore[import-not-found]  # noqa: E402
+    EnvCfg,
+    EvalCfg,
+    EvalDatasetCfg,
+    SceneCfg,
+    TaskCfg,
+)
+
+eval_cfg = EvalCfg(
+    agent=AgentCfg(
+        # Server port — keep clear of CMA(8087)/N1(8028) so we can run side-by-side.
+        server_port=8096,
+        model_name="ryworld",
+        # Dummy skeleton agent ignores ckpt_path; real wrapper will load this.
+        ckpt_path="/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+        model_settings={
+            "stop_threshold_override": 0.95,
+            "train_config_path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/stage1_discrete.yaml",
+            "train_config_overrides": ["/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/ryworld/configs/overlays/a100_4gpu_discrete.yaml"],
+        },
+    ),
+    env=EnvCfg(
+        env_type="vln_pe",
+        env_settings={
+            "use_fabric": False,
+            "headless": True,
+        },
+    ),
+    task=TaskCfg(
+        task_name="20260513T030719_step0015000_stop0.95_full835_thr095_shard5",
+        task_settings={
+            "env_num": 1,
+            "use_distributed": False,
+            "proc_num": 1,
+            "max_step": 500,
+            "warm_up_step": 500,
+        },
+        scene=SceneCfg(
+            scene_type="kujiale",
+            scene_data_dir="data/scene_data/vlnverse",
+        ),
+        robot_name="h1",
+        robot_flash=True,
+        robot_usd_path="data/Embodiments/vln-pe/h1/h1_vln_pointcloud.usd",
+        camera_resolution=[256, 256],
+        camera_prim_path="torso_link/h1_pano_camera_0",
+    ),
+    dataset=EvalDatasetCfg(
+        dataset_type="kujiale",
+        dataset_settings={
+            "base_data_dir": "data/vlnverse/raw_data/vlnverse/coarse_shard5",
+            "split_data_types": ["val_unseen"],
+            "filter_stairs": False,
+            "instruction_type": "formal",
+        },
+    ),
+)

eval_results/shards/shard5/eval_result.log

@@ -0,0 +1,15 @@
+[split:val_unseen] total 97 data
+############[val_unseen]#############
+TL = 581.9372461389353 / 97 = 5.9994
+NE = 293.45884496488355 / 97 = 3.0253
+FR = 0 / 97 = 0.0%
+StR = 0 / 97 = 0%
+OS = 80.0 / 97 = 82.47%
+SR = 54.0 / 97 = 55.669999999999995%
+SPL = 52.05919405377673 / 97 = 53.669999999999995%
+detail:
+[reach_goal]:54
+[not_reach_goal]:43
+[success]:54
+[fall]:0
+##########################

eval_results/shards/shard5/meta.json

@@ -0,0 +1,161 @@
+{
+  "eval_id": "20260513T030719_step0015000_stop0.95_full835_thr095_shard5",
+  "tag": "full835_thr095_shard5",
+  "timestamp_recorded": "2026-05-13T17:08:00+0800",
+  "ckpt": {
+    "path": "/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/ckpt_step0015000_final.pt",
+    "step": 15000,
+    "size_bytes": 2812658821
+  },
+  "code_state": {
+    "git_hash": "a4fa65370358d6f251e158f03aced7d30f167da6",
+    "branch": "feat/batched-trunk-forward",
+    "dirty": true,
+    "dirty_files": [
+      "docs/eval_ledger.jsonl",
+      "run_train.sh",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD0_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD1_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD2_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD3_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD4_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD5_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD6_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD7_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu0_thr091_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu1_thr092_port8092.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu2_thr093_port8093.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu3_thr094_port8094.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu4_thr095_port8095.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu5_thr096_port8096.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu6_thr097_port8097.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_FULL_gpu7_thr098_port8098.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v7_port8091.py",
+      "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen_v8_port8092.py",
+      "scripts/eval/run_eval_structured.sh",
+      "third_party/vlnverse_emr",
+      "yworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_coarse_val_unseen.py"
+    ],
+    "dirty_patch_md5": "aa7101e2b529b2f5596267cd1e8073ee"
+  },
+  "eval_config": {
+    "framework": "vlnverse_emr",
+    "base_cfg": "ryworld/eval/vlnverse_plugin/configs/h1_ryworld_cfg_vlnverse_SHARD5_port8096.py",
+    "derived_cfg": "checkpoints/ryworld_stage1_discrete_20260510_025317_cw07_iw25/evals/20260513T030719_step0015000_stop0.95_full835_thr095_shard5/eval_cfg.py",
+    "stop_threshold": 0.95,
+    "task_name": "20260513T030719_step0015000_stop0.95_full835_thr095_shard5"
+  },
+  "data_state": {
+    "scenes_used": [
+      "kujiale_0020",
+      "kujiale_0063",
+      "kujiale_0066",
+      "kujiale_0090",
+      "kujiale_0290"
+    ],
+    "scenes_with_hdr": [
+      "kujiale_0020",
+      "kujiale_0063",
+      "kujiale_0066",
+      "kujiale_0090",
+      "kujiale_0290"
+    ],
+    "scenes_missing_hdr": []
+  },
+  "hardware": {
+    "host": "di-20260502152021-n5hs7",
+    "gpus": [
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      },
+      {
+        "name": "NVIDIA GeForce RTX 4090 D",
+        "mem_mb": 24564
+      }
+    ],
+    "count": 8
+  },
+  "results_aggregate": {
+    "TL": 5.9994,
+    "NE": 3.0253,
+    "FR": 0.0,
+    "StR": 0.0,
+    "OS": 0.8247,
+    "SR": 0.5567,
+    "SPL": 0.5367,
+    "n_reach_goal": 54,
+    "n_not_reach_goal": 43,
+    "n_success": 54,
+    "n_fall": 0,
+    "n_total": 97
+  },
+  "results_per_scene": {
+    "kujiale_0290": {
+      "n": 23,
+      "SR": 0.34782608695652173,
+      "avg_NE": 3.781735910187727,
+      "avg_SPL": 0.3250137721273278
+    },
+    "kujiale_0090": {
+      "n": 24,
+      "SR": 0.4166666666666667,
+      "avg_NE": 3.2836376880803937,
+      "avg_SPL": 0.4008321211523211
+    },
+    "kujiale_0066": {
+      "n": 8,
+      "SR": 0.375,
+      "avg_NE": 4.2475675281598475,
+      "avg_SPL": 0.3727614434639622
+    },
+    "kujiale_0063": {
+      "n": 24,
+      "SR": 0.75,
+      "avg_NE": 2.1600414548636446,
+      "avg_SPL": 0.7393609358769763
+    },
+    "kujiale_0020": {
+      "n": 17,
+      "SR": 0.8235294117647058,
+      "avg_NE": 2.3881783564580426,
+      "avg_SPL": 0.7915322324304835
+    }
+  },
+  "stop_head_stats": {
+    "n_samples": 14980,
+    "stop_prob_min": 0.222,
+    "stop_prob_p50": 0.749,
+    "stop_prob_p90": 0.901,
+    "stop_prob_max": 0.975,
+    "stop_prob_mean": 0.7475921228304419,
+    "fire_rate_pathA": 0.02610146862483311,
+    "fire_rate_pathB": 0.00814419225634179,
+    "no_stop_rate": 0.9657543391188251
+  },
+  "n_episodes_recorded": 96
+}

eval_results/shards/shard5/per_episode.csv

@@ -0,0 +1,97 @@
+episode_id,scene,NE,success,osr,SPL
+kujiale_0290_27_0,kujiale_0290,0.21074272373707986,1,1.0,1.0
+kujiale_0290_29_4,kujiale_0290,3.8475800910607085,0,1.0,0.0
+kujiale_0290_36_1,kujiale_0290,8.175447351718805,0,1.0,0.0
+kujiale_0290_2_3,kujiale_0290,3.7167510589402073,0,1.0,0.0
+kujiale_0290_9_1,kujiale_0290,1.7187703404786137,1,1.0,1.0
+kujiale_0290_6_4,kujiale_0290,3.800309754247035,0,1.0,0.0
+kujiale_0290_21_4,kujiale_0290,0.9016440786894129,1,1.0,1.0
+kujiale_0290_7_2,kujiale_0290,0.4893063776539863,1,1.0,0.6719209475665499
+kujiale_0290_22_2,kujiale_0290,10.921065369645657,0,0.0,0.0
+kujiale_0290_46_4,kujiale_0290,2.2383843374353503,1,1.0,0.803395811361989
+kujiale_0290_50_0,kujiale_0290,3.796541262317142,0,1.0,0.0
+kujiale_0290_11_0,kujiale_0290,6.2496536590244105,0,0.0,0.0
+kujiale_0290_33_3,kujiale_0290,5.214787992156439,0,1.0,0.0
+kujiale_0290_35_4,kujiale_0290,8.571959479962729,0,0.0,0.0
+kujiale_0290_44_0,kujiale_0290,4.556250376051046,0,1.0,0.0
+kujiale_0290_24_2,kujiale_0290,3.2344320098536885,0,1.0,0.0
+kujiale_0290_0_0,kujiale_0290,0.4625135998174608,1,1.0,1.0
+kujiale_0290_41_3,kujiale_0290,3.1566917868478415,0,1.0,0.0
+kujiale_0290_53_1,kujiale_0290,3.8809868315403793,0,1.0,0.0
+kujiale_0290_19_4,kujiale_0290,1.1468554237716702,1,1.0,1.0
+kujiale_0290_39_0,kujiale_0290,3.438344722821742,0,1.0,0.0
+kujiale_0290_4_4,kujiale_0290,4.786296934628878,0,1.0,0.0
+kujiale_0290_3_2,kujiale_0290,2.464610371917446,1,1.0,1.0
+kujiale_0090_8_2,kujiale_0090,4.713320886806568,0,0.0,0.0
+kujiale_0090_35_4,kujiale_0090,0.4420070097037416,1,1.0,1.0
+kujiale_0090_12_1,kujiale_0090,2.3467792531715217,1,1.0,1.0
+kujiale_0090_6_3,kujiale_0090,5.712569249565485,0,0.0,0.0
+kujiale_0090_29_1,kujiale_0090,3.7659101958091563,0,1.0,0.0
+kujiale_0090_18_3,kujiale_0090,5.044967361825141,0,1.0,0.0
+kujiale_0090_27_0,kujiale_0090,0.6359434325809113,1,1.0,1.0
+kujiale_0090_33_2,kujiale_0090,4.591953409150986,0,0.0,0.0
+kujiale_0090_24_0,kujiale_0090,0.9023015303381454,1,1.0,0.9616635919866703
+kujiale_0090_16_1,kujiale_0090,1.955813580703905,1,1.0,1.0
+kujiale_0090_1_3,kujiale_0090,4.929727791447769,0,0.0,0.0
+kujiale_0090_1_2,kujiale_0090,3.8456905902984864,0,1.0,0.0
+kujiale_0090_40_0,kujiale_0090,4.977264918051201,0,0.0,0.0
+kujiale_0090_32_2,kujiale_0090,2.0049257231916258,1,1.0,0.7490361041116385
+kujiale_0090_2_3,kujiale_0090,0.9757069243828542,1,1.0,0.9092712115573972
+kujiale_0090_25_0,kujiale_0090,2.473872462857331,1,1.0,1.0
+kujiale_0090_23_4,kujiale_0090,3.648728288865427,0,1.0,0.0
+kujiale_0090_28_0,kujiale_0090,4.139430902532957,0,0.0,0.0
+kujiale_0090_5_0,kujiale_0090,2.931382378135006,1,1.0,1.0
+kujiale_0090_36_1,kujiale_0090,5.610302018133688,0,0.0,0.0
+kujiale_0090_17_3,kujiale_0090,3.508996634037388,0,1.0,0.0
+kujiale_0090_35_1,kujiale_0090,0.27080634742374526,1,1.0,1.0
+kujiale_0090_43_3,kujiale_0090,4.155516735172876,0,1.0,0.0
+kujiale_0090_38_2,kujiale_0090,5.223386889743524,0,1.0,0.0
+kujiale_0066_5_2,kujiale_0066,6.328855664916546,0,0.0,0.0
+kujiale_0066_23_4,kujiale_0066,6.812827919032394,0,0.0,0.0
+kujiale_0066_37_0,kujiale_0066,0.3762490518859696,1,1.0,0.9820915477116976
+kujiale_0066_11_2,kujiale_0066,2.7112358504289826,1,1.0,1.0
+kujiale_0066_8_0,kujiale_0066,3.282720679880946,0,1.0,0.0
+kujiale_0066_3_0,kujiale_0066,8.290172400431722,0,1.0,0.0
+kujiale_0066_50_4,kujiale_0066,2.3245550611316537,1,1.0,1.0
+kujiale_0066_45_3,kujiale_0066,3.853923597570566,0,0.0,0.0
+kujiale_0063_15_3,kujiale_0063,0.3028078595411775,1,1.0,1.0
+kujiale_0063_27_0,kujiale_0063,1.3938129709371208,1,1.0,0.9238019903024689
+kujiale_0063_11_2,kujiale_0063,0.9341601076465338,1,1.0,1.0
+kujiale_0063_7_4,kujiale_0063,2.441533401493744,1,1.0,0.9116808941581778
+kujiale_0063_14_0,kujiale_0063,4.5459091908078015,0,1.0,0.0
+kujiale_0063_28_0,kujiale_0063,0.9641251865156325,1,1.0,0.9091795765867834
+kujiale_0063_12_1,kujiale_0063,1.154201687324658,1,1.0,1.0
+kujiale_0063_30_2,kujiale_0063,1.7330018463008086,1,1.0,1.0
+kujiale_0063_16_1,kujiale_0063,1.5028910592217453,1,1.0,1.0
+kujiale_0063_32_2,kujiale_0063,2.4252402584470203,1,1.0,1.0
+kujiale_0063_35_1,kujiale_0063,0.6658288599501118,1,1.0,1.0
+kujiale_0063_10_3,kujiale_0063,3.6028757742495316,0,1.0,0.0
+kujiale_0063_28_1,kujiale_0063,0.9378611428684135,1,1.0,1.0
+kujiale_0063_8_2,kujiale_0063,4.56742776503182,0,1.0,0.0
+kujiale_0063_13_3,kujiale_0063,0.8028442931172827,1,1.0,1.0
+kujiale_0063_18_3,kujiale_0063,0.9094169946113326,1,1.0,1.0
+kujiale_0063_10_4,kujiale_0063,6.607204992717134,0,0.0,0.0
+kujiale_0063_2_2,kujiale_0063,1.3700523154321025,1,1.0,1.0
+kujiale_0063_23_4,kujiale_0063,1.639648671105137,1,1.0,1.0
+kujiale_0063_20_4,kujiale_0063,1.0635487487326667,1,1.0,1.0
+kujiale_0063_5_0,kujiale_0063,2.101382388227922,1,1.0,1.0
+kujiale_0063_29_2,kujiale_0063,1.4333225688454714,1,1.0,1.0
+kujiale_0063_2_1,kujiale_0063,5.68822358816328,0,0.0,0.0
+kujiale_0063_4_4,kujiale_0063,3.05367324543902,0,0.0,0.0
+kujiale_0020_15_4,kujiale_0020,1.5546681034054137,1,1.0,1.0
+kujiale_0020_20_4,kujiale_0020,1.2751847086213033,1,1.0,1.0
+kujiale_0020_42_4,kujiale_0020,1.5241788028683914,1,1.0,1.0
+kujiale_0020_26_0,kujiale_0020,2.910583173545808,1,1.0,1.0
+kujiale_0020_22_4,kujiale_0020,1.8182490718335127,1,1.0,1.0
+kujiale_0020_15_3,kujiale_0020,1.2116844665913584,1,1.0,1.0
+kujiale_0020_1_1,kujiale_0020,1.5630531724654637,1,1.0,1.0
+kujiale_0020_32_2,kujiale_0020,4.719888835343296,0,0.0,0.0
+kujiale_0020_13_3,kujiale_0020,2.898063299553074,1,1.0,1.0
+kujiale_0020_23_4,kujiale_0020,1.599296031791861,1,1.0,1.0
+kujiale_0020_44_3,kujiale_0020,2.044788322747493,1,1.0,0.9649409386069607
+kujiale_0020_37_3,kujiale_0020,2.9199074806468905,1,1.0,0.6707731749216151
+kujiale_0020_43_0,kujiale_0020,1.853855138031314,1,1.0,0.8203338377896431
+kujiale_0020_30_2,kujiale_0020,2.8539372415445117,1,1.0,1.0
+kujiale_0020_11_0,kujiale_0020,3.1301338500493294,0,1.0,0.0
+kujiale_0020_10_3,kujiale_0020,2.7652145440994946,1,1.0,1.0
+kujiale_0020_9_2,kujiale_0020,3.9563458166482066,0,1.0,0.0

eval_results/shards/shard5/server.log.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc1c4100b6d2924874b5b9cb314dec9a62701f152a1a509d8f26b3c97fd45958
+size 69146

eval_results/shards/shard5/server_resume.log

@@ -0,0 +1,1319 @@
+WARNING:  Current configuration will not reload as not all conditions are met, please refer to documentation.
+INFO:     Started server process [2406095]
+INFO:     Waiting for application startup.
+INFO:     Application startup complete.
+INFO:     Uvicorn running on http://localhost:8096 (Press CTRL+C to quit)
+You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`.
+PROJECT_ROOT_PATH:/imotion-vepfs/imotion_zxnet/workspace_wei/ry-dynamics-vln-ryworld/third_party/vlnverse_emr
+INFO:     ::1:38022 - "POST /agent/init HTTP/1.1" 201 Created
+STOP_DEBUG disc_argmax=1 stop_prob=0.750 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.757 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.758 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.760 thr=0.95 final=1 no-stop
+INFO:     ::1:35022 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.760 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.761 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.757 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.760 thr=0.95 final=1 no-stop
+INFO:     ::1:35034 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.772 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.782 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.776 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.788 thr=0.95 final=1 no-stop
+INFO:     ::1:35044 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.782 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.791 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.783 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.791 thr=0.95 final=1 no-stop
+INFO:     ::1:35060 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.784 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.792 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.791 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.799 thr=0.95 final=1 no-stop
+INFO:     ::1:35064 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.788 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.799 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.797 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.801 thr=0.95 final=1 no-stop
+INFO:     ::1:35076 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.756 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.773 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.764 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.760 thr=0.95 final=1 no-stop
+INFO:     ::1:35080 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.758 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.767 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.760 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.756 thr=0.95 final=3 no-stop
+INFO:     ::1:35088 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.750 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.762 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.755 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.755 thr=0.95 final=1 no-stop
+INFO:     ::1:35090 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.731 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.736 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.727 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.730 thr=0.95 final=3 no-stop
+INFO:     ::1:35104 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.652 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.659 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.642 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.647 thr=0.95 final=3 no-stop
+INFO:     ::1:33228 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.610 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.617 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.605 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.617 thr=0.95 final=1 no-stop
+INFO:     ::1:33244 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.590 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.603 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.606 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.615 thr=0.95 final=1 no-stop
+INFO:     ::1:33256 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.609 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.626 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.638 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.642 thr=0.95 final=1 no-stop
+INFO:     ::1:33270 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.663 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.680 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.684 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.686 thr=0.95 final=1 no-stop
+INFO:     ::1:33276 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.725 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.733 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.738 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.745 thr=0.95 final=1 no-stop
+INFO:     ::1:33286 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.689 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.699 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.708 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.714 thr=0.95 final=1 no-stop
+INFO:     ::1:33302 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.717 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.720 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.725 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.729 thr=0.95 final=1 no-stop
+INFO:     ::1:33314 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.729 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.736 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.741 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.754 thr=0.95 final=1 no-stop
+INFO:     ::1:33324 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.749 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.758 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.773 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.781 thr=0.95 final=1 no-stop
+INFO:     ::1:33332 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.735 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.744 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.761 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.773 thr=0.95 final=1 no-stop
+INFO:     ::1:33340 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.744 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.754 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.774 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.789 thr=0.95 final=1 no-stop
+INFO:     ::1:33350 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.771 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.782 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.804 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.817 thr=0.95 final=1 no-stop
+INFO:     ::1:33364 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.808 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.818 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.843 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.854 thr=0.95 final=1 no-stop
+INFO:     ::1:33376 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.842 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.853 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.870 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.880 thr=0.95 final=1 no-stop
+INFO:     ::1:53812 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.856 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.869 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.882 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.892 thr=0.95 final=1 no-stop
+INFO:     ::1:53820 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.858 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.872 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.882 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.891 thr=0.95 final=1 no-stop
+INFO:     ::1:53826 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.861 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.872 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.880 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.891 thr=0.95 final=1 no-stop
+INFO:     ::1:53832 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.863 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.875 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.883 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.894 thr=0.95 final=1 no-stop
+INFO:     ::1:53840 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.872 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.884 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.892 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.905 thr=0.95 final=1 no-stop
+INFO:     ::1:53852 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.884 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.895 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.903 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.913 thr=0.95 final=1 no-stop
+INFO:     ::1:53866 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.893 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.902 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.908 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.914 thr=0.95 final=1 no-stop
+INFO:     ::1:53868 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.908 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.914 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.919 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.922 thr=0.95 final=1 no-stop
+INFO:     ::1:53876 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.917 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.923 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.930 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.933 thr=0.95 final=1 no-stop
+INFO:     ::1:53882 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.942 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.945 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.947 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.937 thr=0.95 final=1 no-stop
+INFO:     ::1:53890 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.951 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=1 stop_prob=0.952 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=1 stop_prob=0.942 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.873 thr=0.95 final=0 A-fire
+INFO:     ::1:53900 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+INFO:     ::1:43816 - "POST /agent/ryworld/reset HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.643 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.655 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.658 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.667 thr=0.95 final=2 no-stop
+INFO:     ::1:43826 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.657 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.666 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.672 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.671 thr=0.95 final=1 no-stop
+INFO:     ::1:43836 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.638 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.646 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.656 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.655 thr=0.95 final=1 no-stop
+INFO:     ::1:43852 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.644 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.653 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.666 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.669 thr=0.95 final=1 no-stop
+INFO:     ::1:43866 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.652 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.665 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.674 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.679 thr=0.95 final=1 no-stop
+INFO:     ::1:43878 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.670 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.686 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.700 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.704 thr=0.95 final=3 no-stop
+INFO:     ::1:43890 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.692 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.706 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.717 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.725 thr=0.95 final=3 no-stop
+INFO:     ::1:43902 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.694 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.702 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.715 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.718 thr=0.95 final=3 no-stop
+INFO:     ::1:43904 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.723 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.732 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.742 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.748 thr=0.95 final=3 no-stop
+INFO:     ::1:43918 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.741 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.748 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.756 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.763 thr=0.95 final=3 no-stop
+INFO:     ::1:45564 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.752 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.763 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.769 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.777 thr=0.95 final=1 no-stop
+INFO:     ::1:45574 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.733 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.743 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.754 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.763 thr=0.95 final=1 no-stop
+INFO:     ::1:45576 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.743 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.755 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.764 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.773 thr=0.95 final=1 no-stop
+INFO:     ::1:45590 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.728 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.738 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.749 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.760 thr=0.95 final=1 no-stop
+INFO:     ::1:45594 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.751 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.761 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.775 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.788 thr=0.95 final=1 no-stop
+INFO:     ::1:45608 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.755 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.766 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.778 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.786 thr=0.95 final=1 no-stop
+INFO:     ::1:45618 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.756 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.767 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.782 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.788 thr=0.95 final=1 no-stop
+INFO:     ::1:45620 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.754 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.768 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.784 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.792 thr=0.95 final=1 no-stop
+INFO:     ::1:45626 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.753 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.770 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.786 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.791 thr=0.95 final=1 no-stop
+INFO:     ::1:45640 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.763 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.778 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.794 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.799 thr=0.95 final=1 no-stop
+INFO:     ::1:45654 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.781 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.798 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.815 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.824 thr=0.95 final=1 no-stop
+INFO:     ::1:45670 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.811 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.823 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.842 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.853 thr=0.95 final=1 no-stop
+INFO:     ::1:45678 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.819 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.837 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.854 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.863 thr=0.95 final=1 no-stop
+INFO:     ::1:55210 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.825 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.845 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.862 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.871 thr=0.95 final=1 no-stop
+INFO:     ::1:55220 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.841 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.861 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.874 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.883 thr=0.95 final=1 no-stop
+INFO:     ::1:55234 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.856 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.875 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.888 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.894 thr=0.95 final=1 no-stop
+INFO:     ::1:55248 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.859 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.878 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.889 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.894 thr=0.95 final=1 no-stop
+INFO:     ::1:55264 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.871 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.887 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.897 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.899 thr=0.95 final=1 no-stop
+INFO:     ::1:55266 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.877 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.890 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.900 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.901 thr=0.95 final=1 no-stop
+INFO:     ::1:55272 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.888 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.898 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.911 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.911 thr=0.95 final=1 no-stop
+INFO:     ::1:55282 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.891 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.902 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.913 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.912 thr=0.95 final=1 no-stop
+INFO:     ::1:55286 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.913 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.923 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.933 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.931 thr=0.95 final=1 no-stop
+INFO:     ::1:55288 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.929 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.942 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.946 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.941 thr=0.95 final=1 no-stop
+INFO:     ::1:55290 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.933 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.946 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.951 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=1 stop_prob=0.945 thr=0.95 final=1 no-stop
+INFO:     ::1:55298 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.949 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.955 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=1 stop_prob=0.953 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.932 thr=0.95 final=0 A-fire
+INFO:     ::1:55302 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.943 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.939 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.932 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.888 thr=0.95 final=0 A-fire
+INFO:     ::1:60630 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.942 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.872 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.838 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.758 thr=0.95 final=0 A-fire
+INFO:     ::1:60642 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.941 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.894 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.870 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.799 thr=0.95 final=0 A-fire
+INFO:     ::1:60646 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.953 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.936 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.775 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.694 thr=0.95 final=0 A-fire
+INFO:     ::1:60658 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+INFO:     ::1:40552 - "POST /agent/ryworld/reset HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.479 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.477 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.488 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.479 thr=0.95 final=2 no-stop
+INFO:     ::1:40558 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.465 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.471 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.481 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.462 thr=0.95 final=2 no-stop
+INFO:     ::1:40562 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.445 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.446 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.457 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.449 thr=0.95 final=1 no-stop
+INFO:     ::1:40574 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.465 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.473 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.474 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.470 thr=0.95 final=1 no-stop
+INFO:     ::1:40588 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.465 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.467 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.479 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.475 thr=0.95 final=1 no-stop
+INFO:     ::1:60744 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.472 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.482 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.489 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.485 thr=0.95 final=1 no-stop
+INFO:     ::1:60760 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.475 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.484 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.497 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.496 thr=0.95 final=1 no-stop
+INFO:     ::1:60768 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.504 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.518 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.528 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.529 thr=0.95 final=1 no-stop
+INFO:     ::1:60782 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.548 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.559 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.564 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.564 thr=0.95 final=3 no-stop
+INFO:     ::1:60798 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.580 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.592 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.588 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.590 thr=0.95 final=1 no-stop
+INFO:     ::1:60802 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.613 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.620 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.622 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.626 thr=0.95 final=3 no-stop
+INFO:     ::1:60812 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.588 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.598 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.600 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.599 thr=0.95 final=3 no-stop
+INFO:     ::1:60818 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.602 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.614 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.613 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.613 thr=0.95 final=1 no-stop
+INFO:     ::1:60832 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.605 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.620 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.618 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.616 thr=0.95 final=1 no-stop
+INFO:     ::1:60840 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.603 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.621 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.621 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.621 thr=0.95 final=1 no-stop
+INFO:     ::1:60848 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.585 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.601 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.603 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.602 thr=0.95 final=1 no-stop
+INFO:     ::1:60864 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.657 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.676 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.685 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.688 thr=0.95 final=1 no-stop
+INFO:     ::1:51040 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.680 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.700 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.706 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.714 thr=0.95 final=1 no-stop
+INFO:     ::1:51044 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.662 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.682 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.691 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.696 thr=0.95 final=1 no-stop
+INFO:     ::1:51052 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.691 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.713 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.718 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.722 thr=0.95 final=1 no-stop
+INFO:     ::1:51064 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.721 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.736 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.742 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.749 thr=0.95 final=1 no-stop
+INFO:     ::1:51074 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.736 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.748 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.752 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.756 thr=0.95 final=1 no-stop
+INFO:     ::1:51082 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.739 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.749 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.750 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.756 thr=0.95 final=1 no-stop
+INFO:     ::1:51096 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.717 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.730 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.730 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.741 thr=0.95 final=1 no-stop
+INFO:     ::1:51102 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.727 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.735 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.734 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.741 thr=0.95 final=1 no-stop
+INFO:     ::1:51104 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.670 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.684 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.687 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.698 thr=0.95 final=1 no-stop
+INFO:     ::1:51116 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.624 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.644 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.645 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.652 thr=0.95 final=1 no-stop
+INFO:     ::1:51126 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.599 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.611 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.619 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.626 thr=0.95 final=1 no-stop
+INFO:     ::1:51134 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.558 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.567 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.585 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.596 thr=0.95 final=1 no-stop
+INFO:     ::1:37038 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.571 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.577 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.596 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.608 thr=0.95 final=1 no-stop
+INFO:     ::1:37042 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.576 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.580 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.594 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.613 thr=0.95 final=1 no-stop
+INFO:     ::1:37058 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.613 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.612 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.624 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.646 thr=0.95 final=1 no-stop
+INFO:     ::1:37060 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.654 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.657 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.667 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.680 thr=0.95 final=1 no-stop
+INFO:     ::1:37064 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.689 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.689 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.698 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.714 thr=0.95 final=1 no-stop
+INFO:     ::1:37068 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.724 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.726 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.736 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.750 thr=0.95 final=1 no-stop
+INFO:     ::1:37074 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.764 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.771 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.780 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.795 thr=0.95 final=1 no-stop
+INFO:     ::1:37082 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.784 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.788 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.795 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.806 thr=0.95 final=1 no-stop
+INFO:     ::1:37096 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.783 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.791 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.798 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.814 thr=0.95 final=1 no-stop
+INFO:     ::1:37110 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.811 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.820 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.833 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.847 thr=0.95 final=1 no-stop
+INFO:     ::1:37112 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.819 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.823 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.836 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.846 thr=0.95 final=1 no-stop
+INFO:     ::1:37118 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.826 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.835 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.848 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.853 thr=0.95 final=1 no-stop
+INFO:     ::1:54964 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.848 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.856 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.865 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.872 thr=0.95 final=1 no-stop
+INFO:     ::1:54968 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.861 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.873 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.885 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.885 thr=0.95 final=1 no-stop
+INFO:     ::1:54976 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.884 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.893 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.901 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.901 thr=0.95 final=1 no-stop
+INFO:     ::1:54980 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.908 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.907 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.905 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.890 thr=0.95 final=1 no-stop
+INFO:     ::1:54982 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.921 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.921 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.923 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.895 thr=0.95 final=1 no-stop
+INFO:     ::1:54994 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.918 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.919 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.924 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.900 thr=0.95 final=1 no-stop
+INFO:     ::1:55002 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.925 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.922 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.914 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.849 thr=0.95 final=0 A-fire
+INFO:     ::1:55012 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.932 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.922 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.921 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.871 thr=0.95 final=1 no-stop
+INFO:     ::1:55020 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.963 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.923 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.747 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.677 thr=0.95 final=0 A-fire
+INFO:     ::1:55026 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+INFO:     ::1:59680 - "POST /agent/ryworld/reset HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.674 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.680 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.688 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.685 thr=0.95 final=1 no-stop
+INFO:     ::1:59684 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.685 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.692 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.702 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.698 thr=0.95 final=1 no-stop
+INFO:     ::1:59696 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.700 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.708 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.720 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.718 thr=0.95 final=1 no-stop
+INFO:     ::1:59710 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.700 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.706 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.715 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.716 thr=0.95 final=1 no-stop
+INFO:     ::1:33950 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.723 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.726 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.737 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.741 thr=0.95 final=1 no-stop
+INFO:     ::1:33956 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.737 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.741 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.752 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.755 thr=0.95 final=1 no-stop
+INFO:     ::1:33968 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.758 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.763 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.778 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.780 thr=0.95 final=1 no-stop
+INFO:     ::1:33974 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.770 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.779 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.792 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.797 thr=0.95 final=1 no-stop
+INFO:     ::1:33978 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.777 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.787 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.798 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.803 thr=0.95 final=1 no-stop
+INFO:     ::1:33984 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.796 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.805 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.816 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.821 thr=0.95 final=1 no-stop
+INFO:     ::1:34000 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.806 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.818 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.830 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.832 thr=0.95 final=1 no-stop
+INFO:     ::1:34014 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.796 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.808 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.821 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.825 thr=0.95 final=1 no-stop
+INFO:     ::1:34022 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.799 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.813 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.826 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.831 thr=0.95 final=1 no-stop
+INFO:     ::1:34036 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.818 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.832 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.848 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.853 thr=0.95 final=1 no-stop
+INFO:     ::1:34046 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.828 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.843 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.856 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.862 thr=0.95 final=1 no-stop
+INFO:     ::1:34048 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.858 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.871 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.883 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.886 thr=0.95 final=3 no-stop
+INFO:     ::1:53584 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.885 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.889 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.898 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.902 thr=0.95 final=3 no-stop
+INFO:     ::1:53594 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.898 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.902 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.908 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.913 thr=0.95 final=3 no-stop
+INFO:     ::1:53598 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.913 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.918 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.919 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.922 thr=0.95 final=1 no-stop
+INFO:     ::1:53606 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.911 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.887 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.886 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.850 thr=0.95 final=1 no-stop
+INFO:     ::1:53608 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.905 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.909 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.902 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.867 thr=0.95 final=1 no-stop
+INFO:     ::1:53610 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.919 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.916 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.921 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.882 thr=0.95 final=1 no-stop
+INFO:     ::1:53626 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.925 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.935 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.902 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.835 thr=0.95 final=0 A-fire
+INFO:     ::1:53634 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.937 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.828 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.788 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.735 thr=0.95 final=0 A-fire
+INFO:     ::1:53636 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.951 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.956 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.723 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.654 thr=0.95 final=0 A-fire
+INFO:     ::1:53638 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+INFO:     ::1:59376 - "POST /agent/ryworld/reset HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.564 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.575 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.583 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.595 thr=0.95 final=1 no-stop
+INFO:     ::1:59382 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.576 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.589 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.599 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.608 thr=0.95 final=1 no-stop
+INFO:     ::1:38942 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.598 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.606 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.626 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.644 thr=0.95 final=3 no-stop
+INFO:     ::1:38950 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.631 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.643 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.662 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.676 thr=0.95 final=3 no-stop
+INFO:     ::1:38956 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.661 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.673 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.691 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.701 thr=0.95 final=3 no-stop
+INFO:     ::1:38970 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.686 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.696 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.711 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.716 thr=0.95 final=3 no-stop
+INFO:     ::1:38982 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.659 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.665 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.680 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.689 thr=0.95 final=3 no-stop
+INFO:     ::1:38992 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.684 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.691 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.710 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.717 thr=0.95 final=3 no-stop
+INFO:     ::1:39004 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.689 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.695 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.712 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.714 thr=0.95 final=3 no-stop
+INFO:     ::1:39018 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.685 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.696 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.720 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.731 thr=0.95 final=3 no-stop
+INFO:     ::1:39030 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.680 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.685 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.705 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.715 thr=0.95 final=3 no-stop
+INFO:     ::1:39038 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.671 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.677 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.703 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.710 thr=0.95 final=3 no-stop
+INFO:     ::1:39048 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.661 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.666 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.694 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.705 thr=0.95 final=1 no-stop
+INFO:     ::1:39060 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.661 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.673 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.708 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.715 thr=0.95 final=1 no-stop
+INFO:     ::1:41532 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.680 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.693 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.726 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.738 thr=0.95 final=1 no-stop
+INFO:     ::1:41544 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.722 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.741 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.777 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.782 thr=0.95 final=1 no-stop
+INFO:     ::1:41556 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.745 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.762 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.801 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.804 thr=0.95 final=1 no-stop
+INFO:     ::1:41570 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.789 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.811 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.854 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.863 thr=0.95 final=1 no-stop
+INFO:     ::1:41584 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.828 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.850 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.878 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.885 thr=0.95 final=1 no-stop
+INFO:     ::1:41592 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.851 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.870 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.891 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.897 thr=0.95 final=1 no-stop
+INFO:     ::1:41596 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.878 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.891 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.911 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.916 thr=0.95 final=1 no-stop
+INFO:     ::1:41612 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.896 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.913 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.929 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.930 thr=0.95 final=1 no-stop
+INFO:     ::1:41628 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.916 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.927 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.938 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.937 thr=0.95 final=1 no-stop
+INFO:     ::1:41638 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.925 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.935 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.942 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.938 thr=0.95 final=1 no-stop
+INFO:     ::1:41648 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.921 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.931 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.932 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.929 thr=0.95 final=3 no-stop
+INFO:     ::1:41650 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.928 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.929 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.930 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.925 thr=0.95 final=3 no-stop
+INFO:     ::1:57482 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.943 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.944 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.894 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.860 thr=0.95 final=3 no-stop
+INFO:     ::1:57496 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.934 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.900 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.918 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.903 thr=0.95 final=3 no-stop
+INFO:     ::1:57506 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.925 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.931 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.920 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.899 thr=0.95 final=3 no-stop
+INFO:     ::1:57518 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.930 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.898 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.900 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.861 thr=0.95 final=3 no-stop
+INFO:     ::1:57534 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.928 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.934 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.905 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.864 thr=0.95 final=3 no-stop
+INFO:     ::1:57544 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.919 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.923 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.931 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.926 thr=0.95 final=1 no-stop
+INFO:     ::1:57560 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.942 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.945 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.940 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.909 thr=0.95 final=1 no-stop
+INFO:     ::1:57568 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.941 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.943 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.938 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.903 thr=0.95 final=1 no-stop
+INFO:     ::1:57574 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.964 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=3 stop_prob=0.963 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.894 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.795 thr=0.95 final=0 A-fire
+INFO:     ::1:57576 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+INFO:     ::1:58380 - "POST /agent/ryworld/reset HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.668 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.678 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.687 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.673 thr=0.95 final=2 no-stop
+INFO:     ::1:58394 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.685 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.685 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.683 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.681 thr=0.95 final=2 no-stop
+INFO:     ::1:58402 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.674 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.679 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.697 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.689 thr=0.95 final=2 no-stop
+INFO:     ::1:58408 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.668 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.675 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.686 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.686 thr=0.95 final=2 no-stop
+INFO:     ::1:58422 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.672 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.676 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.683 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.680 thr=0.95 final=1 no-stop
+INFO:     ::1:58432 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.672 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.674 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.685 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.692 thr=0.95 final=1 no-stop
+INFO:     ::1:38738 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.696 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.700 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.711 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.714 thr=0.95 final=1 no-stop
+INFO:     ::1:38750 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.672 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.678 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.683 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.687 thr=0.95 final=2 no-stop
+INFO:     ::1:38754 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.703 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.711 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.718 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.724 thr=0.95 final=2 no-stop
+INFO:     ::1:38768 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.722 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.731 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.740 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.749 thr=0.95 final=1 no-stop
+INFO:     ::1:38776 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.740 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.750 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.755 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.761 thr=0.95 final=1 no-stop
+INFO:     ::1:38790 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.759 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.777 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.789 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.795 thr=0.95 final=1 no-stop
+INFO:     ::1:38800 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.759 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.780 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.782 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.781 thr=0.95 final=1 no-stop
+INFO:     ::1:38812 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.760 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.775 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.780 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.778 thr=0.95 final=1 no-stop
+INFO:     ::1:38824 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.791 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.803 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.812 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.817 thr=0.95 final=1 no-stop
+INFO:     ::1:38832 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.792 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.814 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.824 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.824 thr=0.95 final=1 no-stop
+INFO:     ::1:38848 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.803 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.826 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.841 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.846 thr=0.95 final=1 no-stop
+INFO:     ::1:38852 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.815 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.835 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.850 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.853 thr=0.95 final=1 no-stop
+INFO:     ::1:38860 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.818 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.833 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.847 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.847 thr=0.95 final=1 no-stop
+INFO:     ::1:59636 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.831 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.839 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.852 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.851 thr=0.95 final=1 no-stop
+INFO:     ::1:59644 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.841 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.848 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.858 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.862 thr=0.95 final=1 no-stop
+INFO:     ::1:59660 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.823 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.837 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.849 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.856 thr=0.95 final=1 no-stop
+INFO:     ::1:59664 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.846 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.855 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.865 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.871 thr=0.95 final=1 no-stop
+INFO:     ::1:59674 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.843 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.857 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.878 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.884 thr=0.95 final=1 no-stop
+INFO:     ::1:59686 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.844 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.857 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.869 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.874 thr=0.95 final=1 no-stop
+INFO:     ::1:59690 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.798 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.812 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.822 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.830 thr=0.95 final=1 no-stop
+INFO:     ::1:59706 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.786 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.794 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.807 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.816 thr=0.95 final=1 no-stop
+INFO:     ::1:59710 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.857 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.867 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.875 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.876 thr=0.95 final=2 no-stop
+INFO:     ::1:59724 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.874 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.880 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.881 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.887 thr=0.95 final=1 no-stop
+INFO:     ::1:59732 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.897 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.911 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.870 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.838 thr=0.95 final=0 A-fire
+INFO:     ::1:47000 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.939 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.795 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.755 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.674 thr=0.95 final=0 A-fire
+INFO:     ::1:47008 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.941 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.948 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.742 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.669 thr=0.95 final=0 A-fire
+INFO:     ::1:47014 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.958 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.752 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.708 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.635 thr=0.95 final=0 A-fire
+INFO:     ::1:47028 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+INFO:     ::1:37960 - "POST /agent/ryworld/reset HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.571 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.582 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.595 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=2 stop_prob=0.594 thr=0.95 final=2 no-stop
+INFO:     ::1:37966 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.581 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.590 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.583 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.580 thr=0.95 final=1 no-stop
+INFO:     ::1:37978 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.590 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.600 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.601 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.596 thr=0.95 final=1 no-stop
+INFO:     ::1:37980 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.597 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.608 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.611 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.610 thr=0.95 final=3 no-stop
+INFO:     ::1:37994 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=2 stop_prob=0.603 thr=0.95 final=2 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.611 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.604 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.610 thr=0.95 final=3 no-stop
+INFO:     ::1:38008 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.625 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.630 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.627 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.632 thr=0.95 final=3 no-stop
+INFO:     ::1:38016 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.642 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.645 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.646 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.649 thr=0.95 final=3 no-stop
+INFO:     ::1:38018 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.637 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.641 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.640 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.645 thr=0.95 final=3 no-stop
+INFO:     ::1:38030 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.640 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.646 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.643 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.651 thr=0.95 final=3 no-stop
+INFO:     ::1:55502 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.664 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.669 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.668 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.677 thr=0.95 final=1 no-stop
+INFO:     ::1:55504 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.642 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.649 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.648 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.657 thr=0.95 final=1 no-stop
+INFO:     ::1:55518 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.657 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.663 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.665 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.672 thr=0.95 final=1 no-stop
+INFO:     ::1:55534 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.654 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.663 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.674 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.681 thr=0.95 final=1 no-stop
+INFO:     ::1:55548 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.671 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.683 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.692 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.700 thr=0.95 final=1 no-stop
+INFO:     ::1:55552 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.684 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.693 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.704 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.709 thr=0.95 final=1 no-stop
+INFO:     ::1:55562 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.685 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.692 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.696 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.701 thr=0.95 final=1 no-stop
+INFO:     ::1:55568 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.689 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.698 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.700 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.706 thr=0.95 final=1 no-stop
+INFO:     ::1:55572 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.703 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.711 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.717 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.722 thr=0.95 final=1 no-stop
+INFO:     ::1:55584 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.711 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.720 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.730 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.735 thr=0.95 final=1 no-stop
+INFO:     ::1:55598 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.722 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.732 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.741 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.746 thr=0.95 final=1 no-stop
+INFO:     ::1:55604 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.740 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.750 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.759 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.766 thr=0.95 final=1 no-stop
+INFO:     ::1:55612 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.737 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.748 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.759 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.764 thr=0.95 final=1 no-stop
+INFO:     ::1:42446 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.771 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.784 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.798 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.807 thr=0.95 final=1 no-stop
+INFO:     ::1:42458 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.785 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.797 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.810 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.814 thr=0.95 final=1 no-stop
+INFO:     ::1:42470 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.802 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.813 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.829 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.826 thr=0.95 final=1 no-stop
+INFO:     ::1:42480 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.821 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.831 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.845 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.839 thr=0.95 final=1 no-stop
+INFO:     ::1:42494 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.839 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.851 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.864 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.863 thr=0.95 final=1 no-stop
+INFO:     ::1:42506 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.854 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.865 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.873 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.869 thr=0.95 final=1 no-stop
+INFO:     ::1:42518 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.866 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.878 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.890 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.886 thr=0.95 final=1 no-stop
+INFO:     ::1:42530 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.878 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.893 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.902 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.897 thr=0.95 final=1 no-stop
+INFO:     ::1:42538 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.893 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.910 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.915 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.912 thr=0.95 final=1 no-stop
+INFO:     ::1:42542 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.897 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.911 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.920 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.917 thr=0.95 final=1 no-stop
+INFO:     ::1:42550 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.889 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.905 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.911 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.907 thr=0.95 final=1 no-stop
+INFO:     ::1:42554 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.884 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.899 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.903 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.897 thr=0.95 final=1 no-stop
+INFO:     ::1:42556 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.881 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.893 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.897 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.897 thr=0.95 final=1 no-stop
+INFO:     ::1:52654 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.874 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.889 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.895 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.895 thr=0.95 final=1 no-stop
+INFO:     ::1:52660 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.867 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.884 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.886 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.881 thr=0.95 final=1 no-stop
+INFO:     ::1:52668 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.879 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.890 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.892 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.886 thr=0.95 final=3 no-stop
+INFO:     ::1:52678 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.890 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=1 stop_prob=0.897 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.900 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.897 thr=0.95 final=3 no-stop
+INFO:     ::1:52690 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.883 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.887 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.894 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.887 thr=0.95 final=3 no-stop
+INFO:     ::1:52702 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=3 stop_prob=0.883 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.885 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.889 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.884 thr=0.95 final=3 no-stop
+INFO:     ::1:52716 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=1 stop_prob=0.926 thr=0.95 final=1 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.934 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=3 stop_prob=0.840 thr=0.95 final=3 no-stop
+STOP_DEBUG disc_argmax=0 stop_prob=0.782 thr=0.95 final=0 A-fire
+INFO:     ::1:52728 - "POST /agent/ryworld/step HTTP/1.1" 200 OK
+STOP_DEBUG disc_argmax=0 stop_prob=0.955 thr=0.95 final=0 B-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.718 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.685 thr=0.95 final=0 A-fire
+STOP_DEBUG disc_argmax=0 stop_prob=0.609 thr=0.95 final=0 A-fire
+INFO:     ::1:52744 - "POST /agent/ryworld/step HTTP/1.1" 200 OK