GT-111/agibotworld_alpha_v3_eef

AgiBotWorld-Alpha (LeRobot v3 EE)

This is a LeRobot v3 format conversion of the AgiBotWorld-Alpha dataset, reorganized into a canonical bimanual 16D end-effector (EE) pose representation for multi-dataset robot learning. 30 Hz, 33,609 episodes, ~56.9M frames.

Original Dataset

AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems

Bu, Q., Cai, J., Chen, L., et al. (AgiBot-World-Contributors). AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems. arXiv:2503.06669, 2025.

• Original Source: https://huggingface.co/datasets/agibot-world/AgiBotWorld-Alpha
• Project Website: https://agibot-world.com
• GitHub: https://github.com/OpenDriveLab/AgiBot-World

AgiBot World is a large-scale real-world robot manipulation dataset. The Alpha release contains over 1 million trajectories across 217 tasks in five deployment scenarios, featuring dual-arm manipulation with both gripper and dexterous hand embodiments.

License

This dataset is released under CC BY-NC-SA 4.0, consistent with the original AgiBotWorld-Alpha license.

Important: By downloading this dataset, you agree to the AgiBot World Community License Agreement. The original dataset is gated — please also request access at the original Hugging Face repo.

Conversion Details

What we changed

1. Canonical Bimanual EE Pose Representation: The raw per-joint proprioception (joint positions, end-effector positions, quaternions, gripper values) is merged into a unified 16D bimanual EE pose vector — per arm: position xyz (metres) + quaternion in scalar-first WXYZ order + gripper, left arm first then right arm.

2. Next-Observation-Absolute Actions: Actions are stored as 16D absolute bimanual EE pose commands with the same per-arm layout. Row t stores the pose that the next observation row t+1 attains (next-observation-absolute alignment).

3. LeRobot v3 Format: Converted to the LeRobot v3 dataset layout with sharded video storage (MP4, AV1 codec) and Parquet-based frame data at 30 Hz.

What we preserved

• All 3 camera views: observation.images.head, observation.images.wrist_left, observation.images.wrist_right (480x640 RGB video)
• Episode structure and task labels (1,057 task variants)
• Frame-level timestamps

Observation and action convention

• observation.state: 16D bimanual end-effector pose — per arm [x, y, z, qw, qx, qy, qz, gripper], left arm (dims 0-7) then right arm (dims 8-15).
• action: 16D absolute bimanual end-effector pose command with the same layout; row t is the pose attained by observation row t+1.
• transition_action_valid is true exactly on rows whose action is a valid next-observation pose; terminal placeholder rows repeat the current observation state and are masked false.
• Gripper channels are raw position, higher = more open.

Dataset Structure

agibotworld_v3_eef/
├── data/
│   └── chunk-*/file-*.parquet            # Frame data (16D state + 16D action + task labels)
├── videos/
│   ├── observation.images.head/chunk-*/file-*.mp4
│   ├── observation.images.wrist_left/chunk-*/file-*.mp4
│   └── observation.images.wrist_right/chunk-*/file-*.mp4
├── meta/
│   ├── info.json                         # Dataset metadata (features, fps, splits)
│   ├── tasks.parquet                     # Task vocabulary
│   ├── episodes/chunk-*/file-*.parquet   # Episode index
│   └── stats.json                        # Per-feature dataset statistics
├── norm_stats.json                       # Post-pipeline normalization stats
└── README.md

Statistics

Metric	Value
Total Episodes	33,609
Total Frames	56,858,070
Total Tasks	1,057
FPS	30
Robot Type	A2D (AgiBot dual-arm)
Embodiment	Gripper
State Dim	16
Action Dim	16

Usage

from lerobot.datasets import LeRobotDataset

dataset = LeRobotDataset("GT-111/agibotworld_v3_eef")

Citation

If you use this dataset, please cite the original AgiBot World paper:

@article{agibotworld2025,
  title={AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems},
  author={AgiBot-World-Contributors and Bu, Qingwen and Cai, Jisong and Chen, Li and Cui, Xiuqi and Ding, Yan and Feng, Siyuan and Gao, Shenyuan and He, Xindong and Hu, Xuan and Huang, Xu and Jiang, Shu and Jiang, Yuxin and Jing, Cheng and Li, Hongyang and Li, Jialu and Liu, Chiming and Liu, Yi and Lu, Yuxiang and others},
  journal={arXiv preprint arXiv:2503.06669},
  year={2025}
}

Version History

• v3.0 (current): LeRobot v3 conversion with canonical 16D bimanual EE pose layout and next-observation-absolute actions

Canonical action contract v2

All child datasets use eef_absolute_next_observation_wxyz_v2. State/action rows are absolute metres + WXYZ quaternion + gripper openness; a valid action at t is the next persisted observation at t+1. Native timestamps and FPS are retained. Terminal, broken, demonstration, or dimension-masked targets contribute to no action, future, progress, or normalization loss. See each meta/action_contract.json and the root collection_manifest.json.