OM-1 (Omnibody Model 1)

Overview

OM-1 (Omnibody Model 1) is Reward AI's first general-purpose robot manipulation policy. Where most learned manipulation policies are trained on teleoperated or on-robot trajectories, OM-1 is trained exclusively on human demonstrations recorded through the Omnibody Hand, a wearable 7-DoF device that captures tactile, proximity, in-hand global-shutter camera and hand-pose data while a person works at full human speed. Reward AI describes the design goal as "One Model, One Data Interface, Any Body": all manipulation data flows through a single interface into a single model, with no split between pre-training and fine-tuning and no on-robot data-collection round before deployment. A separate high-frequency control layer, trained with reinforcement learning in simulation, executes the model's actions on the target robot. The vendor states the model spans robot arms, legged humanoids and wheeled mobile manipulators, covers whole-body work as well as tabletop manipulation, and picks up a new task from less than 30 minutes of human data. OM-1 was announced on 2026-09-14 when the company came out of stealth. As of 2026-09-19 it is an in-house policy: no weights, code, dataset or API have been released, and no success rates, public-baseline comparisons or paper have been published.