chaoqi-liu/mshab_settable_clean_w_depth

---
configs:
- config_name: default
data_files: data/*/*.parquet
tags:
- lerobot
- LeRobot
- robotics
- mshab
- ms-hab
- settable
- simulation
task_categories:
- robotics
pretty_name: mshab_settable_clean_w_depth
---

# mshab_settable_clean_w_depth

LeRobot-format conversion of the released MS-HAB SetTable subtasks.

## Source

- released source repo: `datasets/arth-shukla/MS-HAB-SetTable`
- benchmark scope: SetTable subtasks only
- task strings in this LeRobot dataset: `set table: pick object`, `set table: place object`
- split mapping: LeRobot `train` contains all exported released episodes
- fps: `20`

## Dataset Structure

[meta/info.json](https://huggingface.co/datasets/chaoqi-liu/mshab_settable_clean_w_depth/blob/main/meta/info.json):

```json
{
"codebase_version": "v3.0",
"robot_type": "fetch",
"total_episodes": 4000,
"total_frames": 151547,
"total_tasks": 2,
"fps": 20,
"splits": {"train": "0:4000"},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4"
}
```

- [meta/stats.json](https://huggingface.co/datasets/chaoqi-liu/mshab_settable_clean_w_depth/blob/main/meta/stats.json)
- [meta/tasks.parquet](https://huggingface.co/datasets/chaoqi-liu/mshab_settable_clean_w_depth/blob/main/meta/tasks.parquet)
- [meta/episodes/chunk-000/file-000.parquet](https://huggingface.co/datasets/chaoqi-liu/mshab_settable_clean_w_depth/blob/main/meta/episodes/chunk-000/file-000.parquet)

## Counts

### Per task instruction
- `set table: pick object`: `2000` episodes
- `set table: place object`: `2000` episodes

### Per released source config
- `pick-013_apple`: `1000` episodes
- `pick-024_bowl`: `1000` episodes
- `place-013_apple`: `1000` episodes
- `place-024_bowl`: `1000` episodes

## Observation mapping

- `observation.state`: 42-D flat benchmark state
- `observation.agent.qpos`: `12` dims
- `observation.agent.qvel`: `12` dims
- `observation.extra.tcp_pose_wrt_base`: `7` dims
- `observation.extra.obj_pose_wrt_base`: `7` dims
- `observation.extra.goal_pos_wrt_base`: `3` dims
- `observation.extra.is_grasped`: `1` dims
- `observation.images.fetch_head`: RGB video `(128, 128, 3)`
- `observation.images.fetch_hand`: RGB video `(128, 128, 3)`
- `observation.depth.fetch_head`: int16 depth `(128, 128, 1)`
- `observation.depth.fetch_hand`: int16 depth `(128, 128, 1)`
- `observation.camera.*`: preserved camera matrices from the released HDF5 source

## Action mapping

- `action`: 13-D float32 control in controller order
- names: `arm.shoulder_pan_joint, arm.shoulder_lift_joint, arm.upperarm_roll_joint, arm.elbow_flex_joint, arm.forearm_roll_joint, arm.wrist_flex_joint, arm.wrist_roll_joint, gripper.control, body.head_pan_joint, body.head_tilt_joint, body.torso_lift_joint, base.forward_velocity, base.rotation_velocity`

## Episode metadata

- per-episode released-source metadata is stored in `meta/mshab_source_episodes.jsonl`
- conversion summary and verification are stored in `meta/mshab_settable_export_report.json`

## Notes
- MS-HAB SetTable release stores one observation before each action, so export aligns observation[t] with action[t].
- RGB cameras are stored as LeRobot videos under observation.images.fetch_head and observation.images.fetch_hand.
- Depth streams are preserved as int16 arrays under observation.depth.fetch_head and observation.depth.fetch_hand.
- Task strings are natural-language instructions such as `set table: open drawer`, not short benchmark labels.
- Episodes are truncated at the first successful action frame when success is observed.
- Released HDF5 action rows are clipped to the normalized MS-HAB env action box before export.
- The full 8-way released config identity remains available in meta/mshab_source_episodes.jsonl.
- This variant was derived from the existing LeRobot dataset `datasets/chaoqi-liu/mshab_settable_w_depth` without re-encoding RGB videos.