Reward AI launches OM-1 robot foundation model

Reward AI has introduced OM-1, a robot foundation model the company says can transfer manipulation skills across tabletop arms, industrial arms, mobile manipulators and humanoids without robot-specific training data.

OM-1 is trained from human manipulation demonstrations rather than teleoperation or on-robot experience. Reward AI’s approach pairs the model with a wearable system called the Omnibody Hand, capturing hand motion and interaction directly before mapping that behaviour onto different robot bodies.

The Omnibody Hand is a seven-degree-of-freedom wearable built on the team’s earlier DexCap research. It captures hand pose alongside tactile, proximity and visual data while a person performs tasks at normal speed. Reward AI says it supplements visual-inertial tracking with electromagnetic sensing to reduce error during fast movements.

OM-1 produces higher-level actions including motion direction, speed, force and the timing of grasping and other key events. A separate high-frequency control layer, trained with reinforcement learning in simulation, converts those actions into actuation while accounting for the dynamics, disturbances and delays of each robot body.

The company’s demonstrations include opening a refrigerator, folding laundry, sorting objects, packaging phones with multiple arms, unplugging a latched Ethernet cable and other contact-rich tasks. Reward AI says OM-1 can learn a new task from less than 30 minutes of human demonstration data.

Reward AI has not released broad success-rate benchmarks, model weights, training data, code or a public API, so the cross-embodiment and performance claims remain to be tested outside the launch material.

Have a robotics update Korthos should review? Send news, deployments, product releases, funding rounds, research, or media to tips@korthosrobotics.com or reach out on X at @agkorthos.