Reward AI unveils OM-1 robot manipulation policy
Reward AI has announced OM-1 (Omnibody Model 1), a general-purpose robotic manipulation policy engineered to master dexterous physical tasks without using teleoperation or robot-specific training data. Instead of tethering data collection to slow, expensive robot hardware rigs, OM-1 trains on human demonstrations recorded using a lightweight sensorized wearable hand device. Operating on a "One Model, One Data Interface, Any Body" design philosophy, OM-1 decouples manipulation data scaling from physical robot availability, enabling human task demonstrations to transfer across varied embodiments including robotic arms, tabletop manipulators, and humanoids.
Hardware-bound teleoperation has become the single biggest bottleneck in robotics, and Reward AI's human-wearable collection pipeline could dramatically accelerate data scaling if cross-embodiment retargeting proves robust. Bypassing on-robot teleoperation enables faster, cheaper, and more natural demonstration gathering without needing fleets of physical robots. Hardware-agnostic data interfaces mean models can be trained today for robot embodiments that have not even been manufactured yet. Mapping human kinematic and tactile demonstrations onto disparate mechanical end-effectors remains an open challenge that often degrades precision in edge cases. Currently operating as a proprietary internal policy, OM-1 will need third-party benchmarks and broader access to substantiate its cross-hardware transfer claims.
DISCOVERED
1h ago
2026-09-15
PUBLISHED
2h ago
2026-09-15
RELEVANCE
AUTHOR
luyileo