README.md

---
license: apache-2.0
pretty_name: RoboAug Datasets
task_categories:
  - robotics
tags:
  - robotics
  - manipulation
  - teleoperation
  - data-augmentation
  - hdf5
size_categories:
  - 1K<n<10K
---

# RoboAug Datasets

These datasets accompany the paper **RoboAug: One Annotation to Hundreds of
Scenes via Region-Contrastive Data Augmentation for Robotic Manipulation**
([arXiv:2602.14032](https://arxiv.org/abs/2602.14032),
[project page](https://x-roboaug.github.io),
[code](https://github.com/Open-X-Humanoid/RoboAug)).

This repository contains two kinds of data:

1. **`annotation_images`** — keyframe images from multiple tasks and camera
   views, annotated with bounding boxes in the LabelMe format.
2. **Real-robot manipulation datasets** — teleoperated demonstrations stored as
   per-episode HDF5 files (multi-camera RGB/depth images + robot proprioceptive
   states) on three platforms:

| Dataset | Platform | Tasks | Episodes | Frames | Cameras | Language | Size |
|---------|----------|------:|---------:|-------:|---------|----------|-----:|
| `Single_Arm_UR_5e` | Universal Robots UR5e, single 6-DoF arm + gripper | 5 | 789 | 105,704 | up to 6: front / left / right / top / wrist_left / wrist_right | yes | 149 GB |
| `AgileX_Cobot_Magic_V2.0` | AgileX Cobot Magic V2.0, dual 7-DoF arms | 5 | 254 | 97,359 | camera_top | yes | 8.6 GB |
| `Tien_Kung_2.0` | Tien Kung 2.0 humanoid, dual 7-DoF arms + grippers | 5 | 258 | 45,670 | camera_head | no | 33 GB |

## Download

```bash
pip install -U "huggingface_hub[hf_transfer]"

# Everything
hf download X-Humanoid/RoboAug-Datasets --repo-type dataset --local-dir ./data

# A single subset, e.g. the UR5e demonstrations
hf download X-Humanoid/RoboAug-Datasets --repo-type dataset \
    --include "Single_Arm_UR_5e/*" --local-dir ./data
```

## Layout of the robot datasets

```
<dataset>/
└── <task_name>/
    └── success_episodes/
        └── <episode_id>/                  # e.g. 0307_154247
            └── data/trajectory.hdf5
```

There is no train/val split: every episode of a task sits directly under
`success_episodes/` and appears exactly once. Build your own split if needed.

## Common conventions

- Images are stored as **encoded byte strings** (HDF5 vlen `uint8`), one per
  time step. Color frames are **JPEG** (decode to `(H, W, 3)` uint8); depth
  frames are **PNG** (decode to `(H, W)` uint16, millimeters).
- JPEGs use the standard color order: they display correctly in any image viewer
  and PIL decodes them to RGB. `cv2.imdecode` returns **BGR**, so convert with
  `cv2.cvtColor(img, cv2.COLOR_BGR2RGB)`.
- `master/*` are the leader (teleoperation) signals; `puppet/*` are the follower
  (actually executed) signals — use `puppet/*` as action / state targets.
- Camera resolutions differ across cameras and datasets; read the shape from the
  decoded array.

## 1. `Single_Arm_UR_5e`

| Key | Shape | dtype | Description |
|-----|-------|-------|-------------|
| `language_instruction` | scalar | string | natural-language task instruction |
| `is_intervene` | `(T,)` | int | human-intervention flag (1 = teleoperated frame) |
| `master/arm_joint_position` | `(T, 6)` | float64 | leader 6-axis joint angles |
| `master/hand_joint_position` | `(T, 1)` | float64 | leader gripper |
| `puppet/arm_joint_position` | `(T, 6)` | float32 | follower 6-axis joint angles |
| `puppet/hand_joint_position` | `(T, 1)` | float32 | follower gripper |
| `puppet/end_effector` | `(T, 6)` or `(T, 7)` | float32 | follower TCP pose (6 = position + Euler, 7 = position + quaternion) |
| `observations/rgb_images/<camera>` | `(T,)` | object | JPEG color: `camera_front` / `camera_top` 720x1280, other cameras 480x640 |
| `observations/depth_images/<camera>` | `(T,)` | object | PNG uint16 depth, same resolution as the color stream |

| Task | Episodes | Frames | Cameras |
|------|---------:|-------:|---------|
| ur_move_lemon_from_plate_to_bowl | 295 | 46,766 | front, left, top, wrist_left |
| ur_put_corn_into_the_pot | 294 | 33,725 | front, left, top, wrist_left |
| ur_stack_bowl_250523 | 100 | 11,609 | front, left, right, top, wrist_left, wrist_right |
| ur_open_drawer_and_put_corn | 50 | 7,289 | front, left, right, top, wrist_left, wrist_right |
| ur_put_carrot_and_close_drawer | 50 | 6,315 | front, left, right, top, wrist_left, wrist_right |

> **Known issue:** in `ur_put_carrot_and_close_drawer`, `camera_wrist_left`
> carries no usable signal — every frame of every episode is the same still
> image. Use `camera_wrist_right` (or the other views) for this task.

## 2. `AgileX_Cobot_Magic_V2.0`

Two arms, with each arm's gripper **packed into the joint vector**.

| Key | Shape | dtype | Description |
|-----|-------|-------|-------------|
| `language_instruction` | scalar | string | natural-language task instruction |
| `is_intervene` | `(T,)` | int | human-intervention flag |
| `master/arm_joint_position` | `(T, 14)` | float64 | leader, both arms: 2 x (6 joints + 1 gripper) |
| `puppet/arm_joint_position` | `(T, 14)` | float32 | follower, both arms (same layout) |
| `puppet/end_effector` | `(T, 12)` | float32 | follower TCP pose, both arms: 2 x (position + Euler) |
| `observations/rgb_images/camera_top` | `(T,)` | object | JPEG color, 480x640 |
| `observations/depth_images/camera_top` | `(T,)` | object | PNG uint16 depth, 400x640 |

| Task | Episodes | Frames |
|------|---------:|-------:|
| agilex_1_open_pot_and_put_corn | 50 | 22,276 |
| agilex_1_pick_banana_and_close_drawer | 49 | 17,311 |
| agilex_1_put_corn_into_plate | 55 | 18,497 |
| agilex_1_stack_bowl_250524 | 51 | 18,669 |
| agilex_1_upright_mug | 49 | 20,606 |

## 3. `Tien_Kung_2.0`

Camera data lives under `camera_observations` (`color_images` instead of
`rgb_images`), left/right arms and grippers are separate groups each holding a
`data` dataset, and there is **no** `language_instruction` or end-effector pose.

| Key | Shape | dtype | Description |
|-----|-------|-------|-------------|
| `camera_observations/color_images/camera_head` | `(T,)` | object | JPEG color, 720x1280 |
| `camera_observations/depth_images/camera_head` | `(T,)` | object | PNG uint16 depth, 720x1280 |
| `camera_observations/is_intervene` | `(T,)` | bool | human-intervention flag |
| `camera_observations/timestamp` | `(T,)` | float64 | Unix timestamp per frame |
| `master/arm_{left,right}_position_align/data` | `(T, 7)` | float64 | leader arm 7-DoF joints |
| `master/end_effector_{left,right}_position_align/data` | `(T, 1)` | float64 | leader gripper |
| `puppet/arm_{left,right}_position_align/data` | `(T, 7)` | float32 | follower arm 7-DoF joints |
| `puppet/end_effector_{left,right}_position_align/data` | `(T, 1)` | float32 | follower gripper |

| Task | Episodes | Frames |
|------|---------:|-------:|
| tienkung_16_collect_balls_251016 | 51 | 6,840 |
| tienkung_16_lay_bowl_plate_251031 | 55 | 9,570 |
| tienkung_16_oven | 50 | 8,982 |
| tienkung_16_select_yellow_button | 52 | 9,676 |
| tienkung_16_weight_apple | 50 | 10,602 |

## Reading an episode

```python
import cv2
import h5py
import numpy as np

def decode_rgb(buf):
    return cv2.cvtColor(cv2.imdecode(np.asarray(buf), cv2.IMREAD_COLOR), cv2.COLOR_BGR2RGB)

with h5py.File("Single_Arm_UR_5e/ur_put_corn_into_the_pot/success_episodes/0307_154247/data/trajectory.hdf5", "r") as f:
    instr = f["language_instruction"][()].decode("utf-8")
    rgb = decode_rgb(f["observations/rgb_images/camera_front"][0])       # (720, 1280, 3) RGB
    depth = cv2.imdecode(np.asarray(f["observations/depth_images/camera_front"][0]), cv2.IMREAD_UNCHANGED)
    arm = f["puppet/arm_joint_position"][:]                              # (T, 6)
    hand = f["puppet/hand_joint_position"][:]                            # (T, 1)
```

The [RoboAug repository](https://github.com/Open-X-Humanoid/RoboAug) also ships a
loader for `Single_Arm_UR_5e` (`ReadH5Files` in `src/read_h5.py`); pass
`"to_rgb": True` in its `robot_infor` dict to get RGB arrays.

## License and citation

Released under the **Apache License 2.0**.

```bibtex
@misc{wang2026roboaugannotationhundredsscenes,
      title={RoboAug: One Annotation to Hundreds of Scenes via Region-Contrastive Data Augmentation for Robotic Manipulation},
      author={Xinhua Wang and Kun Wu and Zhen Zhao and Hu Cao and Yinuo Zhao and Zhiyuan Xu and Meng Li and Shichao Fan and Di Wu and Yixue Zhang and Ning Liu and Zhengping Che and Jian Tang},
      year={2026},
      eprint={2602.14032},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2602.14032},
}
```