---
license: odc-by
task_categories:
- robotics
tags:
- robotics
- embodied ai
- manipulation
- mobile manipulation
- humanoid
- unitree-g1
- proprioception
- multi-view video
- LeRobot
language:
- en
size_categories:
- 10M<n<100M
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# FetchMan-data

Simulated training episodes (actions, egocentric video and proprioception) for object pick-up with a
**Unitree G1** humanoid, generated in [MolmoSpaces](https://github.com/allenai/molmospaces).

The dataset was generated from [FetchMan](https://github.com/omarrayyann/fetchman) code developed by Omar Rayyan.

| | |
|---|---|
| Episodes | 352,696 |
| Frames | 52,008,151 |
| Tasks (language instructions) | 902 (`pick up the <object>.`) |
| Control rate | 10 fps |
| Cameras | `observation.head_image`, `observation.wrist_image` (384×224, MP4) |
| Format | [LeRobot](https://github.com/huggingface/lerobot) dataset v3.0 |
| Size | ~510 GB |

## Data layout

```
meta/info.json                # features, fps, path templates
meta/stats.json               # dataset-wide feature statistics
meta/tasks.parquet            # task string -> task_index
meta/episodes/chunk-XXX/*.parquet   # per-episode index (lengths, file offsets, per-episode stats)
data/chunk-XXX/file-XXX.parquet     # per-frame actions and low-dimensional observations
videos/<camera>/chunk-XXX/file-XXX.mp4
```

### Features

| Key | Shape | Description |
|---|---|---|
| `action` | 15 | robot action |
| `observation.joint_pos` | 30 | full-body joint positions |
| `observation.upper_joint_pos` | 11 | upper-body joint positions |
| `observation.right_gripper_pos` | 1 | right gripper position |
| `observation.right_hand_pose` | 7 | right hand pose (position + quaternion) |
| `observation.base_position` / `base_yaw` / `base_height` | 2 / 1 / 1 | base pose |
| `observation.base_rpy` / `base_rp` | 3 / 2 | base orientation |
| `observation.base_velocity` / `base_angular_velocity` | 3 / 3 | base velocities |
| `observation.last_base_vel_cmd` | 3 | last commanded base velocity |
| `observation.target_object_pose` | 7 | pose of the object to pick |
| `observation.target_point` | 2 | target point |
| `observation.head_image`, `observation.wrist_image` | 224×384×3 | camera videos |
| `object_name`, `object_id`, `scene`, `skill_profile` | string | episode metadata |
| `timestamp`, `frame_index`, `episode_index`, `index`, `task_index` | 1 | LeRobot indexing |

See `meta/info.json` for the authoritative feature specification.

## Data access

With LeRobot:

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("allenai/fetchman-data")
frame = ds[0]
```

To download everything (or a subset) locally:

```bash
hf download allenai/fetchman-data --repo-type dataset --local-dir /path/to/fetchman-data
# low-dimensional data only (no videos):
hf download allenai/fetchman-data --repo-type dataset --include "meta/*" "data/*" --local-dir /path/to/fetchman-data
```

## License

This dataset is licensed under [ODC-BY 1.0](https://opendatacommons.org/licenses/by/1-0/).
It is intended for research and educational use in accordance with Ai2's [Responsible Use Guidelines](https://allenai.org/responsible-use).