Datasets:
File size: 3,790 Bytes
5134f66 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 | ---
license: cc-by-4.0
task_categories:
- robotics
tags:
- robotics
- imitation-learning
- bimanual-manipulation
- aloha-agilex
- lerobot
pretty_name: RoboPro
size_categories:
- 10K<n<100K
---
# RoboPro
**RoboPro** is an 80-task bimanual manipulation dataset collected on a dual-arm **Aloha-Agilex** platform. It contains 15,999 expert teleoperation episodes (3.73 M frames at 50 Hz) recorded from three RGB cameras and synchronised 14-DoF joint commands. Each task includes a **clean** variant and ten **cluttered** variants (`d6`…`d15`) with progressively added distractor objects.
The dataset is released in [LeRobot v2.1](https://github.com/huggingface/lerobot) format.
## At a glance
| | |
|---|---|
| Robot | Aloha-Agilex (dual-arm, 14-DoF) |
| Episodes | 15,999 |
| Frames | 3,728,445 |
| FPS | 50 Hz |
| Cameras | 3 × RGB 480×640 H.264 (cam_high, cam_left_wrist, cam_right_wrist) |
| Tasks | 80 base × {clean, d6…d15} variants, 1,622 unique language prompts |
| Format | LeRobot v2.1 (Parquet + MP4) |
| License | CC-BY-4.0 |
## Layout
```
lerobot/roboreal_all_80tasks/
├── meta/
│ ├── info.json # schema, fps, episode count
│ ├── tasks.jsonl # 1,622 task language descriptions
│ ├── episodes.jsonl # per-episode metadata
│ └── episodes_stats.jsonl
├── data/
│ └── chunk-{000..015}/
│ └── episode_{000000..015998}.parquet # 50 Hz proprioception
└── videos/
└── chunk-{000..015}/
└── observation.images.{cam_high,cam_left_wrist,cam_right_wrist}/
└── episode_{000000..015998}.mp4 # H.264 yuv420p, GOP=2
```
Each Parquet row contains:
- `observation.state` — `float32[14]` joint state
- `action` — `float32[14]` target joint command
- `timestamp`, `frame_index`, `episode_index`, `index`, `task_index`
The 14-DoF channel order for both `observation.state` and `action` is:
```
left_waist, left_shoulder, left_elbow, left_forearm_roll,
left_wrist_angle, left_wrist_rotate, left_gripper,
right_waist, right_shoulder, right_elbow, right_forearm_roll,
right_wrist_angle, right_wrist_rotate, right_gripper
```
## Loading
```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset(
repo_id="Hoshipu/RoboPro",
root="lerobot/roboreal_all_80tasks", # subfolder within the repo
)
```
Or stream individual files:
```python
from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq
p = hf_hub_download(
"Hoshipu/RoboPro",
"lerobot/roboreal_all_80tasks/data/chunk-000/episode_000000.parquet",
repo_type="dataset",
)
table = pq.read_table(p)
```
## Croissant
A validated [Croissant 1.0](https://mlcommons.org/croissant) metadata file is included at
`croissant.json` and conforms to the MLCommons Croissant + RAI specification.
## Citation
```bibtex
@dataset{roboPro2026,
title = {RoboPro: 80-Task Bimanual Manipulation Demonstrations on Aloha-Agilex},
author = {Li, Zhiyuan},
year = {2026},
url = {https://huggingface.co/datasets/Hoshipu/RoboPro}
}
```
## Limitations & ethical considerations
- All demonstrations were recorded on a single Aloha-Agilex unit in a fixed laboratory environment; lighting, camera intrinsics, and table geometry do not vary across episodes.
- Tasks are short-horizon (<10 s) tabletop manipulation — no long-horizon, navigation, or contact-rich (insertion, peg-in-hole) tasks.
- Demonstrations reflect the kinematic preferences of a small operator pool; they are not guaranteed to be task-optimal.
- No human faces, voices, biometric or other personal information is captured. Scenes consist of inanimate objects on a laboratory table.
See `croissant.json` for the full RAI block.
|