AIRoA A-0 robot model learns two-armed tasks from 16,151 hours of human handheld demos, no robot data collected
AIRoA, a research team funded by Japan's NEDO, has released A-0, a two-armed robot system trained on 16,151 hours of demonstrations people recorded with YUBI, its open-source handheld gripper. The team says it collected no robot data: no teleoperation, no reinforcement learning, no on-robot corrections. The model starts from the public pi0.5 robot model's weights.
The same gripper and wrist cameras mount on robot arms, so the team says one set of weights runs on Franka, TX-G2, OpenArm and UR robots. On ten tasks, like sorting small parts and folding a towel, A-0 averaged 78% success on a Franka at a site absent from its training data, versus 14% to 27% for four other models post-trained on the same demos.
These are the team's own results in an arXiv report. Its demo clips are autonomous; the Franka clips play at 3x speed. Total pretraining compute isn't reported.
About the footage: Autonomous, sped up: AIRoA's project page labels every robot clip autonomous; Franka and OpenArm clips play at 3x, UR at 2x and TX-G2 at 5x. This clip (Franka) is 3x.
The source post
Post by AIRoA on X.
View the post on X
Sources
- Kris Wu on X
- AIRoA: AIRoA-0 (A-0) project page
- arXiv: AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data
New to robots? Read our guide: How to tell if a robot video is real: CGI, sped up, or teleoperated?.
Summarized by Servospace from the sources above, drafted with AI help and checked against those sources before publishing. Claims are the companies' or outlets' unless we say otherwise. How we work.