---
license: apache-2.0
pretty_name: OOJU Galaxea R1 Pro Pick Cube
task_categories:
  - robotics
size_categories:
  - 10K<n<100K
tags:
  - robotics
  - real-world
  - teleoperation
  - imitation-learning
  - manipulation
  - bimanual
  - galaxea
  - galaxea-r1-pro
  - quest-3
  - meta-quest-3
  - hand-tracking
  - retargeting
  - embodied-ai
---

# OOJU Galaxea R1 Pro Pick Cube

**177 bimanual cube-grasping demonstrations for the Galaxea R1 Pro, captured with
Meta Quest 3 hand tracking and retargeted to the robot.** Each episode carries
full 26-joint hand poses for both hands, first-person passthrough camera frames,
and 22-DoF robot joint positions, velocities and drive targets sampled at a
nominal 50 Hz. The set totals 38,027 XR/camera frames and 80,014 robot samples
across two grasp strategies, side grasp and top grasp.

No teleoperation rig was used. The demonstrations were recorded by a person
wearing a headset and moving their hands, then converted to robot joint commands.

## OOJU dataset series

| Dataset | Robot | Episodes | Robot actions | Format |
| --- | --- | ---: | --- | --- |
| [wild-mani-kitchen](https://huggingface.co/datasets/OOJU/wild-mani-kitchen) | retarget targets: Unitree G1, UR5e | 102 | no | LeRobot v3.0 parquet |
| **ooju-galaxea-pick-cube** (this card) | Galaxea R1 Pro | 177 | **yes, 80,014 samples** | raw JSON + TAR |
| [unitree-g1-pick-place-cup](https://huggingface.co/datasets/OOJU/pick-place-cup-20260404) | Unitree G1 | 47 | yes | LeRobot parquet + policy baselines |

If you want published success rates for policies trained on OOJU data, the cup
dataset is the one that ships evaluation results.

## Tasks

| Folder | Description | Episodes |
| --- | --- | ---: |
| `side-grasp/` | Pick up a cube using a side grasp | 120 |
| `top-grasp/` | Pick up a cube using a top grasp | 57 |

## Snapshot

| Property | Value |
| --- | ---: |
| Episodes released | 177 |
| Recorded task time | 1,600.3 s (26.7 min), mean 9.0 s per episode |
| XR/camera frames | 38,027 |
| Robot samples | 80,014 |
| Robot degrees of freedom | 22 |
| Robot nominal sample rate | 50 Hz |
| JPEG frames (inside archives) | 38,027 |
| Frame archives | 177 |
| JSON files | 1,062 |
| Extracted size | ~5.4 GB |
| Collection dates | 2026-08-13 to 2026-08-16 |

The 177 episodes are the set retained after validation. Recordings that failed
the workspace, joint-limit or teleoperation quality checks were not carried into
this release, so this snapshot contains no failed demonstrations and the recorded
task time above is the time in the released episodes, not the total time spent
collecting.

## Recording details

| Property | Value |
| --- | --- |
| Robot embodiment | Galaxea R1 Pro with grippers |
| Recording mode | Bimanual, hand tracking (no teleoperation rig) |
| XR device | Meta Quest 3 |
| XR tracking | Full hand-joint poses, 26 joints per hand |
| Camera | Left passthrough camera (`passthrough_left`) |
| Camera resolution | 1280 x 1280 RGB |
| Observed camera rate | 23.08 to 25.47 FPS |

## Loading

This repository keeps the native recorder output as JSON and JPEG. It is not
converted to LeRobot or another training format, so `load_dataset` will not work.
Pull the files directly:

```python
from huggingface_hub import snapshot_download

path = snapshot_download("OOJU/ooju-galaxea-pick-cube", repo_type="dataset")
```

```bash
# one episode, trajectory + robot stream + camera frames
tar -xf side-grasp/side-grasp_20260815_210713/camera_real/frames.tar -C ./frames
```

<!-- IF WE CONVERT TO PARQUET, REPLACE THE BLOCK ABOVE WITH:
```python
from datasets import load_dataset
ds = load_dataset("OOJU/ooju-galaxea-pick-cube", split="train")
```
and add `lerobot` + `format:parquet` to the tags, and a `configs:` block to the
frontmatter. Without that the Hub viewer stays off and no Croissant record is
generated, so the dataset will not appear in Google Dataset Search.
-->

## What is in each episode

Each recording has a task name and timestamp identifier, for example
`side-grasp_20260815_210713` or `top-grasp_20260813_213502`:

- `<task>_<timestamp>.json` — Time-series XR frames, including hand and eye
  poses and validation signals.
- `<task>_<timestamp>_metadata.json` — Episode and environment metadata,
  including task environment, device, timing, success, and quality checks.
- `<task>_<timestamp>_robot.json` — Robot observations and actions, including
  joint positions, velocities, drive targets, root pose, and engine state.
- `camera_real/frames.tar` — Uncompressed archive of the original JPEG
  passthrough frames (`real_frame_*.jpg`).
- `camera_real/` also contains camera poses, calibration, and video sidecar
  metadata.

### Episode layout

```text
side-grasp/
├── side-grasp_20260815_210713/
│   ├── side-grasp_20260815_210713.json
│   ├── side-grasp_20260815_210713_metadata.json
│   ├── side-grasp_20260815_210713_robot.json
│   └── camera_real/
│       ├── side-grasp_20260815_210713_camera_poses.json
│       ├── side-grasp_20260815_210713_video.json
│       ├── calib.json
│       └── frames.tar
└── ...
```

### Robot joint ordering

The 22-element `q`, `qd`, and `action_q` arrays use this order:

1. Four torso joints
2. Seven left-arm joints
3. Seven right-arm joints
4. Two left-gripper finger joints
5. Two right-gripper finger joints

Exact joint names are recorded in each `*_robot.json` file.

### Hand-joint ordering

Each hand stores 26 poses in this order: palm and wrist; thumb metacarpal,
proximal, distal, tip; then metacarpal, proximal, intermediate, distal and tip
for index, middle, ring and little fingers. The names are repeated in each frame
under `left_hand.joint_names` and `right_hand.joint_names`. Every hand pose is
translation followed by rotation, `[x, y, z, qx, qy, qz, qw]`.

### Units and coordinate conventions

- Joint angles in radians, translations in metres, time in seconds.
- Quaternions are XYZW, scalar last.
- Robot root poses use the Unity left-handed, Y-up world frame.
- Camera poses are camera-to-world transforms in the Unity world frame.
- Camera intrinsics follow the OpenCV pinhole convention, top-left origin,
  downward-positive image Y axis.

## Synchronization

The robot and XR/camera streams run at different rates. Robot samples are
nominally 50 Hz; each JPEG maps one-to-one to an XR frame and a camera-pose
entry at roughly 24 FPS. Align with `timestamp_seconds` in the trajectory file
and `trajectory_timestamp_seconds` in the camera-pose sidecar.

Do not substitute `center_eye_pose`, `left_eye_pose` or `right_eye_pose` for the
physical passthrough-camera pose. The eye fields are display-view poses; the
camera-pose sidecar describes the image sensor at capture time.

## Data quality and limitations

- The first JPEG of every episode is blank. It is written before the passthrough
  stream is live, so start from the second frame.
- Camera sidecars flag 678 frames as `IMG_STATIC`.
- The camera distortion model is unmeasured; distortion coefficients are zero.
- Video sidecars contain recorder-generated MP4 path fields, but this snapshot
  ships the original JPEG frames inside `frames.tar`, not MP4 files.
- The snapshot covers two grasp tasks, one robot embodiment, one XR device, one
  camera viewpoint and a limited physical setting.
- It contains no failed demonstrations, no audio, no depth and no force/torque.
- Capture device identifiers in the JSON have been replaced with a neutral label.
  All episodes come from a single headset.

## What is visible in the frames

The passthrough frames are first person and were recorded indoors on a high
floor. They show the demonstrator's bare hands and forearms, the desk surface and
props used for the task, and a window with a view of the surrounding area. If you
plan to redistribute derived imagery, note that the window view is identifiable.

## FAQ

**Can I collect demonstration data without building a teleoperation rig?**
Yes. Every episode here was recorded by a person wearing a Meta Quest 3 and
moving their hands; the robot joint commands were produced afterwards by
retargeting. There is no leader arm, no exoskeleton and no motion-capture stage.

**Can data recorded from human hands be used on a robot?**
It has to be retargeted first. This repository ships the result of that step for
the Galaxea R1 Pro: 80,014 robot samples with 22-DoF joint positions, velocities
and drive targets, alongside the original hand poses. The retargeting is what
turns a hand trajectory into something a robot controller can execute.

**How many demonstrations do I need?**
This dataset does not answer that on its own. Our
[cup dataset](https://huggingface.co/datasets/OOJU/pick-place-cup-20260404)
ships trained checkpoints and evaluation results for one task at 47
demonstrations, and reports both the learned-policy and the retrieval-and-replay
numbers so the gap is visible.

**What is the difference between this and a teleoperation dataset?**
Most teleoperation datasets record the leader device or the robot's own
commanded state. This records the human hand at 26 joints per hand plus the
first-person view, and derives the robot action from it. The hand pose is kept,
not discarded, so grasp shape survives into the released data.

## Collect and validate your own data

If your task, objects or environment are not covered here, the same pipeline is
available: Quest 3 capture, automated validation against a target embodiment, and
export.

**[Request access](https://ooju-app.world/request?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=galaxea-pick-cube&utm_content=footer)**

## License

Released under the Apache 2.0 License.

## Citation

```bibtex
@misc{ooju_galaxea_pick_cube,
  title={OOJU Galaxea R1 Pro Pick Cube},
  author={OOJU},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/OOJU/ooju-galaxea-pick-cube}}
}
```