Humanoid
AgiBot World
robot manipulation; imitation learning
Real robot manipulation demonstrations with images, proprioception and hierarchical task labels from mobile dual-arm robots; Alpha and Beta editions have different scopes.
Access links
Primary places to inspect, download, read about, or reproduce the dataset.
Quick facts
- Vehicle type
- Humanoid
- Environment
- Indoor
- Data origin
- Real-world
- Sensor count
- 1
- Ground truth
- Not reported
- Calibration
- Parameters Only
- Annotations
- Available
Sensor overview
High-level sensor availability before the detailed sensor records below.
Ground truth
Reference scope, method, rate, coverage, and provenance.
Not reported
Calibration and synchronization
Reported calibration level, reproducibility signals, and supporting notes.
- Level
- Parameters Only
- Processed parameters
- Available
- Raw calibration data
- Not reported
- Calibration targets
- Not reported
Parameters only
Processed calibration parameters are reported, but raw calibration data is not reported.
Notes
Paper and author documentation describe episode camera calibration. Complete raw calibration sessions are unverified.
Known or intentional limitations
Reported constraints and characteristics to check before using the dataset.
Alpha, Colosseo/Beta and later releases differ. The selected paper reports 1,001,552 trajectories and 2,976.4 hours, while the maintained Beta description gives 1,003,672; no unversioned count or duration is asserted.
Camera and end-effector configurations vary by episode. The eight-camera collection rig does not establish eight released streams in every episode, and tactile coverage remains unverified.
Human motion capture drives teleoperation; it is not independent robot-pose or force truth. Images, task labels and proprioceptive state provide different kinds of supervision.
Hugging Face marks Alpha and Beta as gated; data files require the host access flow. Documentation is public, but archives were not accessed and no registration or terms acceptance was performed.
Cameras
Camera 1
- Model
- Not reported
- Setup
- Multi Camera
- Count
- 1
- Effective cameras
- 1
- Modality
- Other
- Resolution
- Not reported
- Rate
- 30 Hz
- Shutter
- Not Reported
- Lens type
- Not Reported
- HFOV
- Not reported
- VFOV
- Not reported
- Notes
- Mixed camera configurations vary by episode and end effector; do not assume eight public image streams in every episode.
IMUs
Not reported
GNSS
Not reported
LiDAR
Not reported
Radar
Not reported
Additional sensors
Other sensor
- Name
- Robot proprioception and actions
- Count
- 1
- Rate
- 30 Hz
- Notes
- Recorded joint/end-effector state and teleoperation action products. Physical encoder/end-effector counts and tactile availability vary by episode.
Annotations
Reported annotation availability and task support.
- Availability
- Available
- Format
- Task/subtask language, keyframes, scene/object/skill annotations and timestamped failure-recovery reasons.
- Class count
- Not reported
Semantic segmentation
Not reported
Instance segmentation
Not reported
Object detection
Not reported
Optical flow
Not reported
Depth ground truth
Not reported
Citation
Citation key and BibTeX kept at the end of the page for reference.
@misc{agibot2025agibotworld,
author = {{AgiBot-World}},
title = {AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems},
year = {2025},
url = {https://arxiv.org/abs/2503.06669v4}
}Dataset corrections
Notice a specific issue after reviewing the page?
Report an inaccuracy