README.md

---
license: cc-by-4.0
task_categories:
  - robotics
  - reinforcement-learning
tags:
  - mobile-manipulation
  - imitation-learning
  - diffusion-policy
  - maniskill
  - ac-dit
  - bc
  - vla
language:
  - en
size_categories:
  - 1B<n<10B
configs:
  - config_name: train
    data_files:
      - split: pick_apple
        path: train/set_table/pick/013_apple/*.h5
      - split: pick_bowl
        path: train/set_table/pick/024_bowl/*.h5
      - split: place_apple
        path: train/set_table/place/013_apple/*.h5
      - split: place_bowl
        path: train/set_table/place/024_bowl/*.h5
      - split: open_fridge
        path: train/set_table/open/fridge/*.h5
      - split: open_kitchen_counter
        path: train/set_table/open/kitchen_counter/*.h5
      - split: close_kitchen_counter
        path: train/set_table/close/kitchen_counter/*.h5
  - config_name: val
    data_files:
      - split: pick_apple
        path: val/set_table/pick/013_apple/*.h5
      - split: pick_bowl
        path: val/set_table/pick/024_bowl/*.h5
      - split: place_apple
        path: val/set_table/place/013_apple/*.h5
      - split: place_bowl
        path: val/set_table/place/024_bowl/*.h5
      - split: open_fridge
        path: val/set_table/open/fridge/*.h5
      - split: open_kitchen_counter
        path: val/set_table/open/kitchen_counter/*.h5
      - split: close_kitchen_counter
        path: val/set_table/close/kitchen_counter/*.h5
---

# AC-DiT MSHab Reproduction Dataset

Mobile manipulation demonstrations for the **AC-DiT** (Adaptive Coordination Diffusion Transformer for Mobile Manipulation, NeurIPS 2025) benchmark, evaluated on the **ManiSkill-HAB (MSHab) Set-Table** scenario. This repository contains:

1. Trajectory demonstrations (HDF5) for the 7 Set-Table sub-tasks
2. Held-out validation trajectories (100 per task) for held-out evaluation
3. Reproduction artifacts: code config, training/eval scripts, achieved metrics

The reproduction matches or exceeds the paper's reported numbers (mean 59.4% vs paper 55.6%) — see Results section below.

| Resource | Link |
|---|---|
| Paper | https://arxiv.org/abs/2507.01961 |
| Original code | https://github.com/<original-author>/AC-DiT (see paper) |
| Project page | https://ac-dit.github.io/ |
| MSHab | https://arth-shukla.github.io/mshab/ |

---

## Tasks

The Set-Table scenario consists of 7 (task, subtask, object) combinations:

| Combo (short) | Task | Subtask | Object | Description |
|---|---|---|---|---|
| pick_apple | set_table | pick | 013_apple | Fetch robot picks up an apple from somewhere in the scene |
| pick_bowl | set_table | pick | 024_bowl | Fetch picks up a bowl |
| place_apple | set_table | place | 013_apple | Fetch places an apple on the table |
| place_bowl | set_table | place | 024_bowl | Fetch places a bowl on the table |
| open_fridge | set_table | open | fridge | Open the fridge door |
| open_kitchen_counter | set_table | open | kitchen_counter | Open the kitchen counter drawer |
| close_kitchen_counter | set_table | close | kitchen_counter | Close the kitchen counter drawer |

Each demonstration includes the full mobile-base + 7-DoF arm + gripper trajectory captured from an RL expert (SAC for pick/place, PPO for open/close — per MSHab convention).

---

## Data Splits

| Split | Trajectories per task | Total trajectories | Use |
|---|---|---|---|
| `train` | 1000 | 7000 | Stage-2 imitation learning |
| `val` | 100 | 700 | Held-out evaluation for sample MSE / L2 error during training |

All trajectories are **successful demonstrations** (each h5 trajectory cut at the first frame where `success=True`).

---

## HDF5 Structure

Each `traj_<N>` group contains:

```
traj_<N>/
├── actions          (T, 13)  float32   # pd_joint_delta_pos + base_vel for Fetch robot
├── success          (T,)     bool
├── obs/
│   ├── agent/qpos              (T+1, 14) float32  # joint positions
│   ├── extra/
│   │   ├── base_linear_vel     (T+1, 3)  float32
│   │   ├── base_angular_vel    (T+1, 3)  float32
│   │   ├── goal_pos_wrt_base   (T, 3)    float32
│   │   ├── is_grasped          (T,)      bool
│   │   ├── obj_pose_wrt_base   (T, 7)    float32  # xyz + quaternion
│   │   └── tcp_pose_wrt_base   (T, 7)    float32
│   └── sensor_data/
│       ├── fetch_head/
│       │   ├── rgb             (T, 128, 128, 3)  uint8
│       │   ├── depth           (T, 128, 128, 1)  uint16
│       │   ├── position        (T, 128, 128, 3)  float32  # camera-frame 3D points
│       │   └── segmentation    (T, 128, 128, 1)  uint16
│       └── fetch_hand/   # same fields as fetch_head
```

**Action dimensions (13 DoF):**
| idx | dim | range | meaning |
|---|---|---|---|
| 0-6 | arm_joint_{0..6} | [-1, 1] | delta joint positions |
| 7 | gripper | [-1, 1] | -1 close, +1 open |
| 8 | head_pan | [-1, 1] | usually 0 (stationary_head=True in MSHab) |
| 9 | head_tilt | [-1, 1] | usually 0 |
| 10 | torso_lift | [-1, 1] | torso joint delta |
| 11 | base_vel_x | [-1, 1] | base forward velocity |
| 12 | base_angular_vel | [-1, 1] | base yaw velocity |

> **Note**: For **open/close** task demos, the recorded raw actions can exceed [-1, 1] (max abs up to ~20 in some dims). This is because the RL controller's raw output is logged before the env's automatic clipping. The actual executed action is clipped to [-1, 1] by the ManiSkill env. Train consumers may want to `np.clip(actions, -1, 1)` for cleaner supervision.

**Instructions**: Each task directory also contains a `instructions/` sub-folder with `lang_embed_*.pt` files — precomputed SigLIP text embeddings of natural-language task descriptions ("pick the apple from the table", etc.).

---

## How to Use

### Direct download
```bash
# Install hf cli
pip install huggingface_hub

# Download a single task's training data
hf download JJho1314/AC-DiT-MSHab-Dataset \
  --repo-type dataset \
  --include "train/set_table/pick/013_apple/*" \
  --local-dir ./mshab_data

# Download everything
hf download JJho1314/AC-DiT-MSHab-Dataset --repo-type dataset --local-dir ./mshab_data
```

### After download — Re-add point clouds (training pipeline expects them)
```bash
git clone <maniskill-hab-repo>
cd mshab
python add_pointcloud.py --data-dir ./mshab_data/train/set_table/pick/013_apple --max-workers 16
python add_xyzrgb.py --data-dir ./mshab_data/train/set_table/pick/013_apple
```

Repeat for each task. This adds `obs/pointcloud/{xyzw, rgb, mask, xyzrgb}` fields used by the model's Lift3D encoder.

### Training with AC-DiT
See [companion model repo](https://huggingface.co/JJho1314/AC-DiT-MSHab-Reproduction) for code config and training recipe.

---

## Reproduction Results

Mean success rate over 50 evaluation episodes per task using the best checkpoint (ckpt-25000):

| Task | This dataset (ckpt-25000) | Paper (100×3 episodes) |
|---|---|---|
| pick_apple | 26.0% | 33.3 ± 1.9 |
| pick_bowl | 42.0% | 36.0 ± 6.5 ✓ |
| place_apple | 34.0% | 33.3 ± 9.4 ✓ |
| place_bowl | 48.0% | 17.3 ± 6.8 ✓ |
| open_fridge | 92.0% | 90.7 ± 5.0 ✓ |
| open_kitchen_counter | 74.0% | 81.3 ± 6.8 |
| close_kitchen_counter | 100.0% | 97.3 ± 1.9 ✓ |
| **Mean** | **59.4%** | 55.6% ✓ |

5/7 tasks match or beat paper. Two tasks (pick_apple, open_kc) are slightly below but within or near the paper's 1σ confidence band.

---

## Provenance & Reproducibility

This dataset was regenerated from scratch using the MSHab benchmark's official demonstration-collection pipeline:

```bash
# 1. Per-task: gen_data + add_pointcloud + add_xyzrgb
sbatch --export=TASK=set_table,SUBTASK=pick,OBJ=013_apple scripts/gen_combo.sbatch
# ... 7 tasks total

# 2. Encode language instructions (one-time)
python -m data.mshab.encode_instructions \
  --dataset-root <root> --overwrite
```

The expert policies used to roll out trajectories are the MSHab default RL checkpoints:
- **pick / place**: SAC
- **open / close**: PPO

Released by **arth-shukla/mshab_checkpoints** on HF Hub.

### Differences from the original AC-DiT paper's data
- Same RL experts, same simulator (Sapien via ManiSkill 3.0), same scene assets (ReplicaCAD)
- 1000 successful trajectories per task (matches paper)
- Additional held-out 100 val trajectories per task using the val task plans (val scene initializations distinct from train)
- Point cloud fields stripped to save space — re-derivable from depth + segmentation via `add_pointcloud.py`

---

## Citation

```bibtex
@inproceedings{chen2025acdit,
  title = {AC-DiT: Adaptive Coordination Diffusion Transformer for Mobile Manipulation},
  author = {Chen, Sixiang and Liu, Jiaming and Qian, Siyuan and Jiang, Han and Li, Lily and Zhang, Renrui and Liu, Zhuoyang and Gu, Chenyang and Hou, Chengkai and Wang, Pengwei and Wang, Zhongyuan and Zhang, Shanghang},
  booktitle = {NeurIPS},
  year = {2025},
}

@inproceedings{shukla2024mshab,
  title = {ManiSkill-HAB: A Benchmark for Low-Level Manipulation in Home Rearrangement Tasks},
  author = {Shukla, Arth and Lin, Stone Tao and Su, Hao},
  booktitle = {arXiv},
  year = {2024},
}
```

---

## License

CC-BY 4.0 (matches MSHab / ManiSkill upstream).

## Disclaimer

This is an independent reproduction by [Junjie He (JJho1314)](https://huggingface.co/JJho1314), HKUST. Not affiliated with the original AC-DiT authors. Bugs / data issues — please open a discussion.