Robot learning dataset
AgiBot World (Beta)
AgiBot's complete dataset (the Beta release) of real-robot manipulation trajectories recorded on AgiBot G1 platforms, described as over one million trajectories from 100 robots. The CC BY-NC-SA 4.0 license bars commercial use.
Records last verified Oct 13, 2026 · Send a correction
Licence & use
CC BY-NC-SA 4.0
Non-commercial use
Licence text ↗Terms apply to the checked release. Read publisher conditions for mixed collections, converted formats and gated access.
- Publisher
- AgiBot
- Released
- Mar 1, 2025
- Published size
- 2,976.4 h · 1,003,672 trajectories · 48.1 TB
- Modalities
- rgb, depth, proprioception
- Formats
- HDF5, MP4, PNG, JSON
- Real / simulated
- Real
- Includes humanoid recordings
- Yes
Size & release notes
GitHub publishes 1,003,672 trajectories and ~43.8T; the Beta card publishes 2,976.4 hours and Hugging Face lists 48.1 TB total repository files. The storage figures differ and their relationship is not explained. Native files include HDF5 states, PNG depth and camera videos; LeRobot conversion scripts are provided. The card says action_text/description annotations are not yet available.
Recorded embodiments
Models disclosing this training dataset
More robot learning datasets
Evidence
Sources
- [s1] AgiBot World repository README ↗
OpenDriveLab / AgiBot · official · accessed Oct 13, 2026
- [s2] AgiBotWorld-Beta dataset card ↗
AgiBot · official · accessed Oct 13, 2026
- [s3] AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems ↗
arXiv · paper · published Mar 9, 2025 · accessed Oct 13, 2026
- [s4] AgiBot World Colosseo official platform description ↗
OpenDriveLab · official · accessed Oct 13, 2026
- [s5] Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International ↗
Creative Commons · official · accessed Oct 13, 2026