FrontierSeptember 14, 2026via MarkTechPost
Reward AI Releases OM-1: A Robot Policy Trained on Human Demonstrations Only, With No Teleoperation or On-Robot Data
Why it matters
A significant step toward generalizable robot policies that sidestep the traditional data-collection bottleneck (teleoperation, on-robot training). OM-1's human-demonstration-only approach and rapid task adaptation (30 min to new skill) could reshape how robotics teams approach policy development, but closed weights limit immediate practitioner impact.
Key signals
- OM-1: general-purpose manipulation policy trained on human demonstrations only
- 7-DoF wearable glove for data capture (no teleoperation, no on-robot data)
- Runs on industrial arms and humanoids at human speed
- Task learning from under 30 minutes of demonstration data
- Electromagnetic hand tracking: 60% lower overshoot than visual-inertial at 67 cm/s
- RL-trained control layer runs on independent clock
- Weights, code, and API not yet public
- Published September 14, 2026
The hook
Reward AI's OM-1 learns manipulation from human demos alone—no sim-to-real, no teleoperation, runs at human speed on any arm.
Reward AI has released OM-1 (Omnibody Model 1), a general-purpose manipulation policy trained entirely on human demonstrations captured with a 7-DoF wearable glove, with no teleoperation or on-robot data. The policy runs on industrial arms and humanoids at human speed, learns a new task from under 3…