Reward AI Releases OM-1: A Robot Policy Trained on Human Demonstrations Only, With No Teleoperation or On-Robot Data
| Source: MarkTechPost
Tags: Reward AI, OM-1, robotics, robot policy, manipulation, DexCap, embodied AI, human demonstration
Reward AI's OM-1 learns robot manipulation solely from humans wearing a 7-DoF sensorized glove — no teleoperation or on-robot training data — achieving 60% lower tracking error at high speeds versus visual-inertial systems, though no weights or API are publicly available.
Details
Reward AI, a robotics startup whose founders' prior work includes DexCap, HumanPlus, and ALOHA, has introduced OM-1 (Omnibody Model 1): a general-purpose manipulation policy trained exclusively on human demonstrations recorded through a wearable glove, with no teleoperation data or on-robot data in the training pipeline. The Omnibody Hand is a 7-DoF wearable that captures pinching, flexion, and coupled finger motion without per-user calibration. It pairs high-frequency tactile sensing, proximity sensing, and global-shutter cameras to cover fast interactions like conveyor-belt sorting. Reward AI's key quantitative result: electromagnetic tracking augmented with disturbance compensation reduced mean overshoot error by 60% at the highest test speed (67 cm/s) compared to visual-inertial tracking. The design philosophy, 'One Model, One Data Interface, Any Body,' is the central claim — that demonstrations recorded today can train robot embodiments that do not yet exist, because the data pipeline is embodiment-agnostic. The policy has been validated on industrial arms and humanoids running at human speed. Important caveat: OM-1 is an in-house proprietary system. No weights, code, dataset, or API have been released. This is a research preview, not a product developers can deploy.