Datasets:

UCBProject
/

baseline_3_v4_collection_assets

@@ -2,7 +2,7 @@
 license: cc-by-4.0
 language:
 - en
-pretty_name: baseline_3 v4 — Sim Collection Assets (RESUME on A6000)
 tags:
 - robotics
 - manipulation
@@ -11,133 +11,123 @@ tags:
 - franka
 - dexycb
 - collection-assets
 size_categories:
 - 1B<n<10B
 ---
-# baseline_3 v4 — sim collection RESUME package
-Everything needed for a partner machine to **resume** the baseline_3 v4 sim
-collection that ran on a dev box (RTX 5090) until the orchestrator was killed
-to free GPU. Resume-aware orchestrator detects what's already done and only
-collects the remaining objects.
-> **Intended target: a fresh A6000 dedicated to sim collection** — separate from
-> the A6000 currently training DP3 on `UCBProject/DP3_DexYCB_training_data`.
-> Install env_isaaclab from scratch on the collection machine; do NOT try to
-> share a conda env with the training side.
-**Source repo (code lives here, NOT in this dataset)**:
-[`stzabl-png/UCB_Project @ gate3-curobo-ik`](https://github.com/stzabl-png/UCB_Project/tree/gate3-curobo-ik)
-— specifically `sim/run_grasp_sim_baseline3_v4.py` and
-`scripts/baseline_3_v4/*.sh`. Both shell scripts in this dataset are convenience
-copies of the repo versions — the canonical ones live in the repo.
 ---
-## What's in this package
 | file | size | purpose |
 |------|------|---------|
-| `episodes_g.tar.gz` | 1.2 GB | 900 retargeted DexYCB hdf5 (collector INPUT). Includes ycb_dex_01–20 minus the no-data ones; specifically the **newly retargeted dex_11 (pitcher) + dex_14 (mug)** missing from previous runs. |
 | `obj_usd_cad_ycb.tar.gz` | 489 MB | YCB CAD USDs (17 files, Z-up tagged) + textures. Includes the **new dex_11/14/19 USDs** generated this round. |
-| `episodes_b3_v4_15obj_3yaw_2026-05-25.tar.gz` | 266 MB | Already-collected sim hdf5 for 6 objects (tuna 06, tomato 04, pudding 07, gelatin 08, marker 18, potted_meat 09). |
-| `v4_full15_results_2026-05-25.txt` | 1.5 KB | Results log with 6 DONE_RETRY entries → orchestrator uses this to skip done objects. |
-| `v4_full15_queue_resume.sh` | 4.6 KB | RESUME-aware orchestrator (2-parallel). Copy of repo `scripts/baseline_3_v4/`. |
-| `v4_chunked_with_retry.sh` | 4.6 KB | Per-object chunk-5+retry wrapper. Copy of repo `scripts/baseline_3_v4/`. |
----
-## Before you start — what to install
-### 1. Conda env: `env_isaaclab` (heavy)
-```bash
-conda create -n env_isaaclab python=3.11 -y
-conda activate env_isaaclab
-# 1.1 IsaacSim 5.1 — install the PIP package (NOT the Omniverse Launcher version
-#     under /home/<user>/isaacsim/). The pip package is what `from isaacsim import
-#     SimulationApp` resolves to. Dev box ref: isaacsim==5.1.0.0.
-#     If a 5.0.0-rc version is already installed, force-upgrade:
-pip install --upgrade --extra-index-url=https://pypi.nvidia.com isaacsim==5.1.0
-# Also accept the EULA when prompted on first run.
-# 1.2 cuRobo 0.8 (pip package is `nvidia-curobo`, NOT `curobo`. NOT 0.7 — the
-#     planner uses MotionPlanner / SceneCfg.create API which 0.7 lacks.)
-#     Dev box ref: nvidia-curobo==0.8.0.post1.dev26
-pip install --extra-index-url=https://pypi.nvidia.com nvidia-curobo
-# 1.3 Other deps used by the collector + retry wrapper
-pip install h5py scipy 'numpy<2' termcolor pyyaml requests usd-core
-```
-Verify the install (both must print without error):
-```bash
-python -c "import isaacsim; from isaacsim import SimulationApp; print('IsaacSim 5.1 OK')"
-python -c "from curobo.motion_planner import MotionPlanner; print('cuRobo 0.8 OK')"
-```
-Note: pip package name is `nvidia-curobo` but the Python import name is
-`curobo` — this is the standard NVIDIA naming pattern.
-### 2. Repo (gate3-curobo-ik branch)
-```bash
-git clone -b gate3-curobo-ik git@github.com:stzabl-png/UCB_Project.git
-cd UCB_Project
-# DP3 is already vendored under third_party/3D-Diffusion-Policy/3D-Diffusion-Policy
-# but you do NOT need to install it for sim collection (only for training).
 ```
-### 3. GPU
-- Tested on A6000 48 GB and RTX 5090 32 GB.
-- 2-parallel collection uses ~10 GB sustained, ~18 GB peak (cuRobo subprocess spikes).
-- Single-parallel (`OBJS` loop run sequentially) if VRAM is tight.
 ---
-## Download + extract (run from `UCB_Project/`)
-```bash
-pip install huggingface_hub
-mkdir -p Baseline1/data output
-# Pull tarballs
-huggingface-cli download UCBProject/baseline_3_v4_collection_assets \
-    --repo-type dataset \
-    --local-dir /tmp/v4_pkg
-# Extract into matching layout (uses --strip-components=0 so paths preserved as-is)
-tar xzf /tmp/v4_pkg/episodes_g.tar.gz                               # → Baseline1/data/episodes_g/
-tar xzf /tmp/v4_pkg/obj_usd_cad_ycb.tar.gz                          # → output/obj_usd_cad/ycb/
-tar xzf /tmp/v4_pkg/episodes_b3_v4_15obj_3yaw_2026-05-25.tar.gz     # → Baseline1/data/episodes_b3_v4_15obj_3yaw_2026-05-25/
-# Place the results log + scripts where orchestrator expects
-cp /tmp/v4_pkg/v4_full15_results_2026-05-25.txt /tmp/
-# Optional — only if your repo doesn't already have these
-# (gate3-curobo-ik DOES have them under scripts/baseline_3_v4/, prefer those):
-# cp /tmp/v4_pkg/v4_*.sh /tmp/
-```
-Verify layout:
 ```bash
-ls Baseline1/data/episodes_g/*.hdf5 | wc -l              # expect 900
-ls output/obj_usd_cad/ycb/*.usd | wc -l                  # expect ~17
-ls Baseline1/data/episodes_b3_v4_15obj_3yaw_2026-05-25/*.hdf5 | wc -l   # expect ~153
-grep DONE_RETRY /tmp/v4_full15_results_2026-05-25.txt | wc -l    # expect 6
 ```
----
-## Run the resume
-**This is the only command you need:**
 ```bash
-cd UCB_Project   # repo root
-# Override the env vars to match YOUR layout (replace paths with yours)
 PROJ=$HOME/UCB_Project \
 PY=$HOME/miniconda3/envs/env_isaaclab/bin/python \
 OUT=Baseline1/data/episodes_b3_v4_15obj_3yaw_2026-05-25 \
@@ -146,103 +136,12 @@ PAR=2 \
   bash scripts/baseline_3_v4/v4_full15_queue_resume.sh
 ```
-The orchestrator will:
-1. Read `RESULTS` and find 6 `DONE_RETRY` entries (06, 04, 07, 08, 18, 09) → skip them
-2. Launch the remaining 9 objects (12, 05, 03, 15, 02, 11, 14, 20, *and 19 which has 0 src ep and will silently emit nothing*) at PAR parallel
-3. Each object: chunk-5 + auto-retry on PhysX-pollution sanity fail
-4. New `DONE_RETRY` lines append to `RESULTS` as objects finish
-5. Final per-object summary appended at the end
-### Choosing `PAR` on A6000
-A6000 has 48 GB VRAM but **half** the CUDA cores of RTX 5090 (where this run
-was validated at PAR=2). cuRobo subprocesses (called per chunk by the
-collector) are compute-bound, not memory-bound; historically PAR=4 caused
-65 % `plan_fail` on RTX 5090 from cuRobo GPU contention. The A6000's fewer
-cores likely makes contention **worse**, not better, even with more VRAM.
-Recommendation:
-| `PAR` | Sustained VRAM | Peak VRAM | Wall time (9 obj) | Risk |
-|-------|----------------|-----------|-------------------|------|
-| 1     | ~4 GB          | ~7 GB     | ~9-12 h           | None |
-| **2** | ~8 GB          | ~14-18 GB | **~5-7 h**        | **Validated, default** |
-| 3     | ~12 GB         | ~21-27 GB | ~4-5 h (if no contention drop) | Unknown — VRAM fine, cuRobo contention untested on A6000 |
-| 4     | ~16 GB         | ~28-36 GB | ~3-4 h (theoretical) | High — 65 % plan_fail on 5090; expect same or worse |
-**Start with PAR=2.** After object #1 finishes, look at its `DONE_RETRY` line
-in `RESULTS` and compare `plan_fail` ratio to dev-box reference (35-40 % is
-normal). If plan_fail is comparable, you can bump `PAR=3` for the rest (the
-orchestrator re-reads PAR on each loop iteration, but bumping mid-run requires
-killing + restarting the orchestrator — it'll resume from where it stopped).
----
-## What to watch for — *specific* gotchas verified for this setup
-These are not theoretical — every one was either a confirmed bug we fixed or an
-explicit non-default the partner needs to know about. Cross-checked against the
-actual code, not just docs.
-| # | Gotcha | Status / Action |
-|---|--------|-----------------|
-| 1 | **Object mass = 0.05 kg** is hardcoded in `sim/run_grasp_sim_baseline3_v4.py` L962. Real per-class mass (0.3–0.6 kg) caused PhysX corruption + DP3 eval SR = 0 % when we tried it earlier. **Do not override.** | ✅ Already baked in, no action |
-| 2 | **`--yaw-aug` is True by default**. Producing orig + yaw90 + yaw180 + yaw270 = 4 attempts per source ep. The retry wrapper assumes this (`ATTEMPTS_PER_EP=4`). If you change one without the other, the case-C retry math (`MARKERS / ATTEMPTS_PER_EP`) breaks. | ✅ Defaults aligned, no action |
-| 3 | **`OUT` env var must match the existing collected dir** (`episodes_b3_v4_15obj_3yaw_2026-05-25`), NOT a fresh date — otherwise resume can't see the 6 done objects' hdf5 and the safety check refuses. | ⚠️ Use the env var shown above |
-| 4 | **`RESULTS` env var must point to the shipped log**. If you let it default to `/tmp/v4_full15_results_<TODAY>.txt`, that file won't exist → orchestrator will treat all 15 objects as "not done" and re-collect the 6. | ⚠️ Use the env var shown above |
-| 5 | **`PROJ` and `PY` env vars** override the hardcoded `/home/accelerator/...` defaults baked into the scripts for the dev box. **Always set these on partner machine.** | ⚠️ Use the env vars shown above |
-| 6 | **`scripts/baseline_3_v4/` lives in repo** (gate3-curobo-ik). The orchestrator launches `bash scripts/baseline_3_v4/v4_chunked_with_retry.sh ...` with a relative path — so it must be run from repo root. `cd UCB_Project` before launching. | ⚠️ Run from repo root |
-| 7 | **PAR=2 default** (overridable via env var). 4-par historically broke on RTX 5090 (65 % plan_fail). On A6000, 3-par is plausibly OK due to extra VRAM, but cuRobo compute contention is the binding constraint and A6000 has 50 % fewer CUDA cores — so 3-par is **untested**, may degrade. See "Choosing PAR on A6000" above. | ⚠️ Default 2; bump to 3 only after watching first object's plan_fail rate |
-| 8 | **dex_19 (large_clamp) has 0 source ep**. DexYCB raw simply doesn't contain large_clamp grasp sessions. The orchestrator will launch ycb_dex_19, find 0 hdf5 in `episodes_g/`, and the collector will silently emit nothing. Not a bug. | ✅ Already documented in script |
-| 9 | **Logs from prior runs in `/tmp/v4_obj_NN_*.out`** would contaminate stats. The orchestrator deletes per-object logs only for objects it's about to (re)launch — so done objects keep their original logs. | ✅ Handled |
-| 10 | **First boot is slow**: each new IsaacSim process takes ~30 sec to load all extensions; you'll see thousands of `Failed to create change watch ... errno=28/No space left on device` errors that are **non-fatal** (inotify watch limit; ignore). | ⚠️ Don't panic |
----
-## Object-by-object expectation (based on dev-box run)
-The first 6 (already done, will be skipped):
-| obj | yield (saved) | notes |
-|---|---|---|
-| 06 tuna | 25 ep | clean (0 PhysX warn) |
-| 04 tomato | 29 | normal |
-| 07 pudding | 31 | normal |
-| 08 gelatin | 22 | normal |
-| 09 potted_meat | 36 | normal |
-| 18 marker | 10 | low yield (thin object) |
-Expect the remaining 9 to behave similarly: ~20-35 ep per object, ~50-60 min each at 2-parallel. Total ~5-7 h for the batch.
-dex_19 will produce 0 ep (no source data). Don't be alarmed by `0 saved` for it.
----
-## After the run
-A final summary block appears at the tail of `RESULTS`:
-```
-=== FINAL PER-OBJ SUMMARY (after resume) ===
-  ycb_dex_06: orig=10  yaw=15  total=25
-  ycb_dex_04: orig=12  yaw=17  total=29
-  ...
-  TOTAL: NNN trajectories
-```
-Send that summary back to the dev box, and we'll handle zarr build + DP3
-training. Do **NOT** delete the collected hdf5 dir until then — it's the
-training input.
----
-## Troubleshooting
-- **`ImportError: ...isaacsim`** → wrong conda env activated. Confirm with `which python` then `python -c "from isaacsim import SimulationApp"`.
-- **GPU OOM during 2-parallel** → drop one object out of `OBJS` list temporarily; or run sequentially by editing the orchestrator's `while [[ $(jobs -rp | wc -l) -ge 2 ]]` line → `-ge 1`.
-- **Every object's `plan_fail` count is huge (>80 %)** → likely cuRobo install issue. Run a single object first and grep its log for cuRobo errors. The collector falls back to legacy synthesize+IK if cuRobo plan fails, but ideally cuRobo should work.
-- **Orchestrator launches all 15 objects (didn't skip)** → `RESULTS` env var didn't get exported properly. Confirm `echo "$RESULTS"` inside the same shell where you launch.
-- **Resume hdf5 saved=0 for everything** → `OUT` env var wrong or directory empty; safety check should have caught this but if you see this pattern, abort and verify paths.
 ---

 license: cc-by-4.0
 language:
 - en
+pretty_name: baseline_3 v4 - Sim Collection Assets (ARCHIVED)
 tags:
 - robotics
 - manipulation
 - franka
 - dexycb
 - collection-assets
+- archived
 size_categories:
 - 1B<n<10B
 ---
+# baseline_3 v4 — sim collection assets
+> **2026-05-26 — STATUS: ARCHIVED / REFERENCE-ONLY.**
+>
+> The plan to run sim collection on the A6000 was **abandoned** this round
+> due to a glibc incompatibility: the A6000 system glibc is 2.31 (Ubuntu
+> 20.04) but IsaacSim 5.1.0 requires glibc 2.35 (Ubuntu 24.04). All sim
+> collection (DexYCB + OakInk) was completed on the dev box (RTX 5090).
+>
+> The contents of this dataset (tarballs of source episodes, USDs, scripts,
+> partial results log) are kept for reference and reproducibility — the
+> code links remain accurate — but the "run the resume on A6000"
+> instructions below are **no longer the current plan**.
+>
+> **For the current state of the data and training pipeline, see:**
+> - DexYCB-sourced training trajectories (162 ep) →
+>   [`UCBProject/DP3_DexYCB_training_data`](https://huggingface.co/datasets/UCBProject/DP3_DexYCB_training_data)
+> - OakInk-sourced training trajectories (207 ep) →
+>   [`UCBProject/DP3_OakInk_training_data`](https://huggingface.co/datasets/UCBProject/DP3_OakInk_training_data)
+>
+> The DexYCB README also documents the new **combined 369-ep training run**
+> on A6000.
 ---
+## What's in this archived package
 | file | size | purpose |
 |------|------|---------|
+| `episodes_g.tar.gz` | 1.2 GB | 900 retargeted DexYCB hdf5 (collector INPUT for the sim collection that ran on the dev box). Includes ycb_dex_01–20 minus the no-data ones; specifically the **newly retargeted dex_11 (pitcher) + dex_14 (mug)** missing from previous runs. |
 | `obj_usd_cad_ycb.tar.gz` | 489 MB | YCB CAD USDs (17 files, Z-up tagged) + textures. Includes the **new dex_11/14/19 USDs** generated this round. |
+| `episodes_b3_v4_15obj_3yaw_2026-05-25.tar.gz` | 266 MB | Already-collected sim hdf5 for 6 objects (tuna 06, tomato 04, pudding 07, gelatin 08, marker 18, potted_meat 09) at the time the partial results log was snapshotted. |
+| `v4_full15_results_2026-05-25.txt` | 1.5 KB | Partial results log with 6 DONE_RETRY entries (point-in-time snapshot before the run continued on the dev box). |
+| `v4_full15_queue_resume.sh` | 4.6 KB | Resume-aware orchestrator (2-parallel). Repo canonical version under `scripts/baseline_3_v4/`. |
+| `v4_chunked_with_retry.sh` | 4.6 KB | Per-object chunk-5+retry wrapper. Repo canonical version under `scripts/baseline_3_v4/`. |
+**Source code (canonical)**:
+[`stzabl-png/UCB_Project @ gate3-curobo-ik`](https://github.com/stzabl-png/UCB_Project/tree/gate3-curobo-ik)
+— `sim/run_grasp_sim_baseline3_v4.py` and `scripts/baseline_3_v4/*.sh`.
+The OakInk-specific orchestrator (added 2026-05-26) lives at
+`scripts/baseline_3_v4/oakink_full89_queue_resume.sh` in the same repo.
+---
+## How sim collection actually ran (2026-05-25 + 2026-05-26)
+All on the dev box (RTX 5090, Ubuntu 24.04, glibc 2.35):
+| Run | Date | Result |
+|------|------|--------|
+| DexYCB resume — 15 obj, 2-parallel chunked-5 | 2026-05-25 | **162 successful trajectories** across 10 objects → `UCBProject/DP3_DexYCB_training_data` |
+| OakInk full — 89 obj, 2-parallel chunked-5 | 2026-05-26 | **207 successful trajectories** across 45 objects (44 obj contributed 0 due to obj geometry / Franka kinematic limits) → `UCBProject/DP3_OakInk_training_data` |
+---
+## Why the A6000 was dropped from the collection pipeline
+Diagnosed 2026-05-25:
+```
+$ python -c "from isaacsim import SimulationApp"
+ImportError: /lib/x86_64-linux-gnu/libc.so.6: version 'GLIBC_2.35' not found
+(required by .../isaacsim-5.1.0/_build/lib/libomni.kit.app.so)
 ```
+A6000 system glibc 2.31 is fixed by the Ubuntu 20.04 base; bumping the
+system libc to 2.35 in-place is not safe. Containerizing IsaacSim would
+work but adds a deployment burden we decided wasn't worth it for sim
+collection that runs on the dev box anyway. The A6000 stays on training.
 ---
+## Original "resume on A6000" instructions (DEPRECATED — kept for reference)
+<details>
+<summary>Click to expand the original A6000 sim-collection guide.
+This was written before the glibc issue was diagnosed; it is no longer the
+plan. Follow the new training instructions in the DexYCB dataset README instead.</summary>
+The instructions below describe how the A6000 sim collection *would* have
+worked if glibc had been compatible. They are kept verbatim for
+reproducibility; do not follow them on the current A6000.
+### env_isaaclab install (requires glibc ≥ 2.35)
 ```bash
+conda create -n env_isaaclab python=3.11 -y
+conda activate env_isaaclab
+pip install --upgrade --extra-index-url=https://pypi.nvidia.com isaacsim==5.1.0
+pip install --extra-index-url=https://pypi.nvidia.com nvidia-curobo
+pip install h5py scipy 'numpy<2' termcolor pyyaml requests usd-core
 ```
+### Download + extract from this archive
+```bash
+cd UCB_Project
+pip install huggingface_hub
+mkdir -p Baseline1/data output
+huggingface-cli download UCBProject/baseline_3_v4_collection_assets \
+    --repo-type dataset --local-dir /tmp/v4_pkg
+tar xzf /tmp/v4_pkg/episodes_g.tar.gz
+tar xzf /tmp/v4_pkg/obj_usd_cad_ycb.tar.gz
+tar xzf /tmp/v4_pkg/episodes_b3_v4_15obj_3yaw_2026-05-25.tar.gz
+cp /tmp/v4_pkg/v4_full15_results_2026-05-25.txt /tmp/
+```
+### Run the resume
 ```bash
+cd UCB_Project
 PROJ=$HOME/UCB_Project \
 PY=$HOME/miniconda3/envs/env_isaaclab/bin/python \
 OUT=Baseline1/data/episodes_b3_v4_15obj_3yaw_2026-05-25 \
   bash scripts/baseline_3_v4/v4_full15_queue_resume.sh
 ```
+For PAR tuning, gotcha list, troubleshooting, and per-object expected yield,
+see the original 10-gotcha table that previously appeared here — the
+canonical version is now in the repo's `Baseline1/RETRAIN_V4_FULL12.md` and
+the inline comments in `scripts/baseline_3_v4/v4_*.sh`.
+</details>
 ---