AgiBot GO-1 and AgiBot World

2025-03-10 · Publisher : AgiBot · Domain : Robotics

AgiBot publishes the technical report for AgiBot World, over one million trajectories across 217 tasks, and GO-1 (Genie Operator-1), a generalist policy built on latent actions.

What changes

  • AgiBot World: 1M+ trajectories, 217 tasks, 5 deployment scenarios.
  • GO-1 uses latent action representations to make better use of data.
  • +30% on average vs Open X-Embodiment pretraining, per the authors.
  • Open dataset; GO-1 open-sourced in September 2025.

### Background On March 10, 2025, AgiBot, a Shanghai-based Chinese humanoid maker, published with OpenDriveLab the technical report of AgiBot World and introduced GO-1 (Genie Operator-1). Data had started to be released in late December 2024. ### Specifications - **AgiBot World**: over one million trajectories across 217 tasks in five deployment scenarios, an order of magnitude more than existing datasets according to the authors, with human-in-the-loop quality checks. - **Hardware**: from grippers to dexterous hands with visuo-tactile sensors. - **GO-1**: a generalist policy built on latent action representations, whose performance scales predictably with data volume. - **Results**: +30% on average over Open X-Embodiment pretraining, above 60% success on complex tasks and +32% over RDT (scores published by the authors). - **Openness**: dataset, tools and model released; GO-1 open-sourced in September 2025. ### Why it matters AgiBot bets on large-scale, standardized data collection verified by humans. The release makes a Chinese player one of the main providers of open manipulation data. ### Key takeaway Massive data and generalist policy go hand in hand: AgiBot World joins Open X-Embodiment among the major open robot manipulation corpora.

Links

Tags : VLA, Jeu de données, Open Source, AgiBot, Chine