AgiBot World

Million-trajectory real-robot manipulation dataset from a fleet of 100+ humanoid robots

License

CC-BY-NC-SA-4.0

Overview

AgiBot World is the large-scale manipulation data platform from AgiBot and OpenDriveLab — over one million real-robot trajectories collected on a fleet of more than 100 dual-arm humanoid robots across 100+ real-world scenarios, from contact-rich manipulation to long-horizon tasks in home, retail, industrial and office settings. An IROS 2025 Best Paper Award finalist published in IEEE T-RO, with the AgiBotWorld-2026 release distributed on Hugging Face in LeRobot-compatible formats.

AgiBot World tackles robot learning's data bottleneck with production-scale collection: standardized dual-arm humanoid platforms with dexterous end-effectors, professional teleoperators, and a quality pipeline covering task design, collection protocols and validation. The corpus spans five target domains (household, retail, industrial, restaurant, office) with camera streams, proprioception and language annotations per episode, including deliberately hard skills — contact-rich manipulation, deformable objects, multi-robot collaboration, long-horizon mobile manipulation. The companion GO-1 foundation policy and evaluation protocols demonstrate scaling behaviour on the data. Successive releases (AgiBotWorld-Alpha/Beta, the 2026 edition) ship on Hugging Face with LeRobot-format conversions, making it one of the largest openly downloadable real-robot manipulation corpora and a standard pretraining source for VLA-class models alongside Open X-Embodiment and DROID. The dataset carries a CC BY-NC-SA 4.0 license — research-friendly, non-commercial.

Why it matters

AgiBot World is one of the largest real-robot manipulation datasets ever released — fleet-scale, quality-controlled, humanoid-embodied data of the kind previously locked inside companies, now openly downloadable and already standard in VLA pretraining mixes (IROS 2025 award finalist, T-RO).

Problem it solves

Manipulation policies starve for diverse, high-quality real-robot data; academic collections top out at tens of thousands of episodes. AgiBot World supplies million-scale, professionally collected, consistently embodied trajectories across realistic scenarios.

Key takeaways

• 1M+ real-robot trajectories from a fleet of 100+ dual-arm humanoid robots
• 100+ real scenarios across home, retail, industrial, restaurant and office domains
• Contact-rich, deformable, multi-robot and long-horizon tasks with language annotations
• LeRobot-compatible releases on Hugging Face (AgiBotWorld-2026)
• IROS 2025 Best Paper Award finalist, published in IEEE T-RO; CC BY-NC-SA 4.0

Who it is for

• VLA and manipulation-policy researchers
• Robot-learning teams pretraining on real data
• Imitation-learning benchmark builders
• Embodied-AI dataset curators

Limitations

• CC BY-NC-SA 4.0 — non-commercial use only
• Single (AgiBot) embodiment family — cross-embodiment transfer is on the user
• Full corpus is a very large download (terabyte scale)