A-0 trains on 16k hours of human demos, zero robot data
One set of weights, four robots, and a dataset collected entirely by people holding a gripper.
Kris Wu released A-0, a robot learning system trained on 16,000 hours of human demonstrations with zero hours of robot data collected. Every hour of it was recorded with YUBI, the team's handheld gripper, which a person carries around and operates by hand instead of driving a robot to gather the data.
it's embodiment-agnostic: the same weights run long-horizon, precise manipulation on four different robots
The pitch is that A-0 is a complete system rather than a model drop. Wu says it spans data processing, pre-training, post-training and the inference harness. The paper is out, and the team is releasing code, weights and the dataset.
The claim to watch
The headline claim is that A-0 is embodiment-agnostic, meaning the same weights run on different robot bodies without retraining for each one. Wu says those weights do long-horizon, precise manipulation on four different robots. Long-horizon here means tasks with many steps in sequence, where a mistake early on wrecks everything after it.
That is the part nobody outside the team has reproduced yet. One set of weights crossing four embodiments is the sort of result that gets picked apart once other labs pull the weights down, and the release makes that possible. Which four robots, and how hard the tasks were, is not spelled out in the post itself. Wu points to a page with more demos and details.
Why the handheld rig matters
Collecting robot data normally means a robot, an operator and time on hardware. A handheld gripper means a person can record demonstrations anywhere, with no robot in the loop, which is how you get to a number like 16,000 hours. The open question has always been whether data gathered by a human hand transfers to a machine hand well enough to be worth it.
Chris Paxton's read on the release was short. He called it "More evidence UMI is the way forward", referring to the handheld gripper approach to data collection. That is one researcher's take on one release, not a verdict.
More evidence UMI is the way forward
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.