A million trajectories from one Chinese fleet
AgiBot and OpenDriveLab opened AgiBot World — over a million manipulation trajectories across 217 tasks — along with the GO-1 policy; policies trained on the set gave an average 30% improvement over those trained on Open X-Embodiment.
Why it matters
This is the first robot data commons at Open X-Embodiment scale built outside the United States, and it was measured against Open X-Embodiment rather than presented on its own terms. The field gained a second reference corpus and a number for what a purpose-built collection buys over an aggregated one.
The dataset holds over 1 million trajectories covering 217 tasks across five deployment scenarios. Genie Operator-1, or GO-1, is a generalist policy built on latent action representations. The paper states that policies pre-trained on AgiBot World achieve an average performance improvement of 30% over those trained on Open X-Embodiment, both in-domain and out-of-distribution, and that GO-1 outperforms the prior RDT approach by 32%. The dataset, tools and models were open-sourced. More widely quoted figures of 2,976.4 hours, 87 skills, 106 scenes and over a hundred collecting robots come from the project repository rather than the abstract, and are not used here.