Reward's OM-1 learns new manipulation tasks from under 30 minutes of human demos
chris_j_paxton · x · 2026-09-15
Robotics teams are converging on learning directly from people: ACT-1 learned long-horizon household tasks from human demos; X Square's TwinDEX trained on a few hundred robot-free episodes to run a 24-step chemistry experiment; and Reward's OM-1 learns manipulation from human demonstrations alone, with the same policy running across industrial arms and humanoids on contact-rich tasks requiring force control and recovery. Reward says OM-1 picks up a new task from less than 30 minutes of demo data — and all three teams use dexterous gloves for high-quality data collection.
More from Embodied
- Verne Robotics bets on on-device inference: robots can't rely on WiFi to run policies — danfei_xu · 2026-09-15
- Robot Foundation Models Like Astra May Be Key to Truly Safe Robots — chris_j_paxton · 2026-09-15
- BAAI's World Action Models use video generation to let robots imagine before acting — stepjamUK · 2026-09-15
- Moonphase Flip: a low-distraction AI to-do screen on the back of your phone — future_coded · 2026-09-15
- AI model paints Golden Gate Bridge with a real robot arm, now reproduced in MuJoCo and open-sourced — victormustar · 2026-09-15
- Reimagine Robotics builds for a future where customers teach robots on the job — notmisha · 2026-09-15