Flash ALOHA learns fast, precise tasks from 50 demos
John Yao's team skips teleoperation and has people guide the robot by hand, with paper, code and hardware posted.
John Yao (@johnhyao) and his team released Flash ALOHA, a system they say learns to autonomously complete dynamic and precise manipulation tasks at near human speed after just 50 demonstrations. Everything is open sourced. Yao posted the announcement thread with teammates @Qi_Wu577, @YihuaiGao, Chelsea Finn and @zipengfu credited.
With just 50 demonstrations, Flash ALOHA learns to autonomously complete dynamic and precise tasks at near human speed.
The data collection is the point
Most robot demonstration data comes from teleoperation, where a human drives the robot remotely. That works, but it is slow, and slow data makes slow policies. The other common route is UMI style handheld grippers, where a person holds a gripper and a camera tracks it.
Flash ALOHA skips both. People guide the robot directly, moving the arms by hand. Chelsea Finn, who is on the team, framed the tradeoff in one line. Teleop gives you slower data, and UMI tracking degrades once you move quickly. Guiding the robot itself gives you fast motion and accurate tracking at the same time, because the robot's own joints are doing the measuring.
Speed as the metric
The headline claim is not success rate, it is speed. Plenty of policies finish a task eventually. Near human speed on dynamic and precise tasks is a different bar, and it is the one the team chose to put in the title. Fifty demonstrations is a small number for that claim, which is part of why it is interesting.
Shown versus claimed
The sources here are the two launch posts. Neither spells out which tasks were run, how many trials, or what the success rates were. What is checkable is that the team posted the paper, the code and the hardware design together, so the speed claim is open to anyone who wants to rebuild the setup and time it. Finn's post links the full release. Nobody outside the team has reported results yet.
Teleop yields slower data, while UMI tracking degrades with higher speeds.
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.