A-0 trains robots from 16k hours of human demos with zero robot data
chris_j_paxton · x · 2026-11-06
AIRoA released A-0, a robot learning system trained entirely on 16,151 hours of human demonstrations recorded with YUBI, an open-source handheld gripper — no teleoperation, no RL, no on-robot corrections.
- 10.65M subtask segments, each with task and subtask instructions
- Embodiment-agnostic: same weights run long-horizon, precise manipulation on Franka, TX-G2, OpenArm and UR
- 78% success on ten Franka tasks at a site with no training data, vs 27% for the strongest of four other models
- Post-training on a new task takes under 4 hours on one node of 8 H200s; action network uses two forward passes per query instead of ten
- Paper, code, weights and dataset all open-sourced
chrisjpaxton says this is more evidence the UMI handheld-data approach is the way forward.
More from Embodied
- NVIDIA launches SimReady Central with 3,500+ OpenUSD robot assets physically calibrated against real hardware — CyberRobooo · 2026-11-06
- Shenlong Wang tours 'May the Force Be with Your Pixels' talk on physical world models — YuXiang_IRVL · 2026-11-06
- Moon Lake previews robotics research: simulation is how to scale the RL era of robots — Aiden_Tech_Ai · 2026-11-06
- Beff Jezos hangs out next to Machina Labs' robotic 'drone printer' — beffjezos · 2026-11-06
- Smart ring team clarifies leaked photo shows a sizing ring, real one has a display — that_anokha_boy · 2026-11-06
- Full Physical AI tutorial: train and deploy your own policy on a low-cost SO-101 robot arm — philfung · 2026-11-06