Datasets:

PI-Dojo
/

EgoManipulation-xEmbody

---
license: cc-by-4.0
language:
  - en
task_categories:
  - video-classification
  - robotics
tags:
  - ego
  - embodied-ai
  - physical-ai
  - robotics
  - multimodal
  - human-demonstration
  - egocentric
  - stereo
  - imu
configs:
  - config_name: default
    data_files:
      - split: sample
        path: "data/episodes.parquet"
  - config_name: trajectories
    data_files:
      - split: sample
        path: "assets/*/trajectory.parquet"
  - config_name: cameras
    data_files:
      - split: sample
        path: "assets/*/cameras.parquet"
---
# EgoManipulation Cross-Embodiment (xEmbody)

EgoManipulation Cross-Embodiment (xEmbody) is a PI Dojo dataset that pairs egocentric human demonstrations with motion retargeted to the Unitree H2 humanoid with ALLEX hands. The xEmbody workflow uses stereo video and IMU recordings to estimate metric depth, SLAM camera poses, and 3D hand poses. Those observations support human motion reconstruction and retargeting of both whole-body motion and dexterous hand motion in simulation.

This release contains six episodes with synchronized RGB, depth previews, hand keypoints, human reconstruction, and robot replay data.

## Episodes

All episodes are exported at 30 fps. Within an episode, frame `i` refers to the same instant in all seven videos, the trajectory and camera tables, the hand keypoints, and the SLAM pose file. Frame indices and timestamps start at zero on the trimmed timeline.

## Dataset Viewer

| Configuration | Row unit | Contents |
| --- | --- | --- |
| `default` | One episode | Task label, seven video columns, playback metadata, and paths to episode and robot assets |
| `trajectories` | One frame | 61 actuator angles in radians, floating-base position in meters, and base orientation as a `wxyz` quaternion |
| `cameras` | One frame | Robot head and recording camera poses and intrinsics |

Each configuration uses the `sample` split. Join frame-level tables on `episode_id`, `frame_index`, and `timestamp_ns`.

| Video column | Contents |
| --- | --- |
| `rgb_video` | Trimmed egocentric RGB |
| `stereo_depth_video` | RGB-aligned, colorized stereo depth preview |
| `hand_keypoints_video` | Hand keypoints projected onto RGB |
| `human_reconstruction_video` | Side view of the reconstructed human motion |
| `h2_overlay_video` | Retargeted H2 + ALLEX overlaid on RGB |
| `h2_left_view_video` | Left simulator view with robot head and SLAM camera markers |
| `h2_right_view_video` | Right simulator view with robot head and SLAM camera markers |

The depth video is a colorized visualization of optical-axis depth. Numerical 3D hand positions are available in the hand keypoint Parquet files.

## Episode assets and coordinate frames

Episode files are under `assets/<episode_id>/`. Each `manifest.json` lists its published files and SHA-256 checksums.

- `hand_keypoints.parquet` contains 42 rows per frame, ordered by frame, hand (left then right), and joint index. The `x_m`, `y_m`, and `z_m` fields are in meters in the recording camera optical frame. The `u_px` and `v_px` fields are pixel coordinates in the RGB image. Missing coordinates are null and marked with `valid=false`.
- `trajectory.parquet` and `cameras.parquet` use an episode-local, gravity-aligned replay world with +Z up. Camera optical axes are +X right, +Y down, and +Z forward. Poses are row-major 4×4 camera-to-world transforms. Intrinsics are row-major 3×3 matrices in pixels.
- `slam_cam_pose.tum` preserves the original SLAM world with timestamps in seconds on the trimmed timeline and quaternions in `xyzw` order. Left-multiply each TUM pose by `replay_world_from_slam_world` from `slam_alignment.json` to recover the corresponding `slam_world_from_optical` pose in the camera table.

## Dataset metadata

The root-level `meta.json` is the machine-readable reference for the dataset. It records the schema version, shared frame and timestamp conventions, H2 + ALLEX actuator order, coordinate frames and units, camera pose and intrinsics conventions, the SLAM alignment convention, and the hand keypoint layout. [Dataset metadata](meta.json)

## Load the data

Request access on the [dataset page](https://huggingface.co/datasets/PI-Dojo/EgoManipulation-xEmbody), then authenticate with a Hugging Face account that has access.

```bash
hf auth login
```

```python
from datasets import Video, load_dataset
repo = "PI-Dojo/EgoManipulation-xEmbody"
episodes = load_dataset(repo, "default", split="sample", streaming=True, token=True)
# Read the episode metadata and video paths without decoding the MP4s.
for name, feature in episodes.features.items():
    if isinstance(feature, Video):
        episodes = episodes.cast_column(name, Video(decode=False))
episode = next(iter(episodes))
print(episode["task_label"], episode["num_frames"], episode["rgb_video"]["path"])
```

## Replay H2 + ALLEX

Download the shared robot assets and the trajectory and camera tables for the episode you want to replay. The example uses episode 1.

```bash
hf download PI-Dojo/EgoManipulation-xEmbody --type dataset \
  --include 'robot/**' \
  --include 'assets/1/trajectory.parquet' \
  --include 'assets/1/cameras.parquet' \
  --local-dir ./xembody
cd ./xembody
python -m pip install mujoco numpy pyarrow
python robot/replay.py --episode 1
```

The script checks the robot head camera poses against MuJoCo forward kinematics. Add `--show` for interactive kinematic playback. The URDF, meshes, joint order, and mimic-joint relationships are included under [robot/](robot/).

## License

The dataset card declares CC BY 4.0. The distributed robot assets include their license files. [Unitree license](robot/LICENSE.unitree.txt) · [ALLEX license](robot/LICENSE.allex.txt)

	---
	license: cc-by-4.0
	language:
	- en
	task_categories:
	- video-classification
	- robotics
	tags:
	- ego
	- embodied-ai
	- physical-ai
	- robotics
	- multimodal
	- human-demonstration
	- egocentric
	- stereo
	- imu
	configs:
	- config_name: default
	data_files:
	- split: sample
	path: "data/episodes.parquet"
	- config_name: trajectories
	data_files:
	- split: sample
	path: "assets/*/trajectory.parquet"
	- config_name: cameras
	data_files:
	- split: sample
	path: "assets/*/cameras.parquet"
	---

	# EgoManipulation Cross-Embodiment (xEmbody)

	EgoManipulation Cross-Embodiment (xEmbody) is a PI Dojo dataset that pairs egocentric human demonstrations with motion retargeted to the Unitree H2 humanoid with ALLEX hands. The xEmbody workflow uses stereo video and IMU recordings to estimate metric depth, SLAM camera poses, and 3D hand poses. Those observations support human motion reconstruction and retargeting of both whole-body motion and dexterous hand motion in simulation.

	This release contains six episodes with synchronized RGB, depth previews, hand keypoints, human reconstruction, and robot replay data.

	## Episodes

	All episodes are exported at 30 fps. Within an episode, frame `i` refers to the same instant in all seven videos, the trajectory and camera tables, the hand keypoints, and the SLAM pose file. Frame indices and timestamps start at zero on the trimmed timeline.

	## Dataset Viewer

	\| Configuration \| Row unit \| Contents \|
	\| --- \| --- \| --- \|
	\| `default` \| One episode \| Task label, seven video columns, playback metadata, and paths to episode and robot assets \|
	\| `trajectories` \| One frame \| 61 actuator angles in radians, floating-base position in meters, and base orientation as a `wxyz` quaternion \|
	\| `cameras` \| One frame \| Robot head and recording camera poses and intrinsics \|

	Each configuration uses the `sample` split. Join frame-level tables on `episode_id`, `frame_index`, and `timestamp_ns`.

	\| Video column \| Contents \|
	\| --- \| --- \|
	\| `rgb_video` \| Trimmed egocentric RGB \|
	\| `stereo_depth_video` \| RGB-aligned, colorized stereo depth preview \|
	\| `hand_keypoints_video` \| Hand keypoints projected onto RGB \|
	\| `human_reconstruction_video` \| Side view of the reconstructed human motion \|
	\| `h2_overlay_video` \| Retargeted H2 + ALLEX overlaid on RGB \|
	\| `h2_left_view_video` \| Left simulator view with robot head and SLAM camera markers \|
	\| `h2_right_view_video` \| Right simulator view with robot head and SLAM camera markers \|

	The depth video is a colorized visualization of optical-axis depth. Numerical 3D hand positions are available in the hand keypoint Parquet files.

	## Episode assets and coordinate frames

	Episode files are under `assets/<episode_id>/`. Each `manifest.json` lists its published files and SHA-256 checksums.

	- `hand_keypoints.parquet` contains 42 rows per frame, ordered by frame, hand (left then right), and joint index. The `x_m`, `y_m`, and `z_m` fields are in meters in the recording camera optical frame. The `u_px` and `v_px` fields are pixel coordinates in the RGB image. Missing coordinates are null and marked with `valid=false`.
	- `trajectory.parquet` and `cameras.parquet` use an episode-local, gravity-aligned replay world with +Z up. Camera optical axes are +X right, +Y down, and +Z forward. Poses are row-major 4×4 camera-to-world transforms. Intrinsics are row-major 3×3 matrices in pixels.
	- `slam_cam_pose.tum` preserves the original SLAM world with timestamps in seconds on the trimmed timeline and quaternions in `xyzw` order. Left-multiply each TUM pose by `replay_world_from_slam_world` from `slam_alignment.json` to recover the corresponding `slam_world_from_optical` pose in the camera table.

	## Dataset metadata

	The root-level `meta.json` is the machine-readable reference for the dataset. It records the schema version, shared frame and timestamp conventions, H2 + ALLEX actuator order, coordinate frames and units, camera pose and intrinsics conventions, the SLAM alignment convention, and the hand keypoint layout. [Dataset metadata](meta.json)

	## Load the data

	Request access on the [dataset page](https://huggingface.co/datasets/PI-Dojo/EgoManipulation-xEmbody), then authenticate with a Hugging Face account that has access.

	```bash
	hf auth login
	```

	```python
	from datasets import Video, load_dataset

	repo = "PI-Dojo/EgoManipulation-xEmbody"
	episodes = load_dataset(repo, "default", split="sample", streaming=True, token=True)

	# Read the episode metadata and video paths without decoding the MP4s.
	for name, feature in episodes.features.items():
	if isinstance(feature, Video):
	episodes = episodes.cast_column(name, Video(decode=False))

	episode = next(iter(episodes))
	print(episode["task_label"], episode["num_frames"], episode["rgb_video"]["path"])
	```

	## Replay H2 + ALLEX

	Download the shared robot assets and the trajectory and camera tables for the episode you want to replay. The example uses episode 1.

	```bash
	hf download PI-Dojo/EgoManipulation-xEmbody --type dataset \
	--include 'robot/**' \
	--include 'assets/1/trajectory.parquet' \
	--include 'assets/1/cameras.parquet' \
	--local-dir ./xembody

	cd ./xembody
	python -m pip install mujoco numpy pyarrow
	python robot/replay.py --episode 1
	```

	The script checks the robot head camera poses against MuJoCo forward kinematics. Add `--show` for interactive kinematic playback. The URDF, meshes, joint order, and mimic-joint relationships are included under [robot/](robot/).

	## License

	The dataset card declares CC BY 4.0. The distributed robot assets include their license files. [Unitree license](robot/LICENSE.unitree.txt) · [ALLEX license](robot/LICENSE.allex.txt)