Reward AI's OM-1 learns manipulation from gloved humans, not robot data

Robotics startup Reward AI has released OM-1, short for Omnibody Model 1, a general-purpose manipulation policy that learns from people wearing a sensorized glove and then runs on industrial arms and humanoids at human speed. The claim the company highlights: no teleoperation data and no on-robot data enter training. The team's prior work includes DexCap, HumanPlus and ALOHA.

OM-1 is currently an in-house policy. No weights, code, dataset or API have been published, so developers cannot run it on their own hardware.

The glove and the data interface

The stack starts with Omnibody Hand, a wearable that builds on the team's earlier DexCap portable motion capture work. Instead of copying the human hand joint by joint, it uses a seven-degree-of-freedom design centred on choosing contact points, reorienting objects in hand, and switching between precision and power grasps. The glove combines high-frequency tactile sensing, proximity sensing for the pre-contact approach, and global-shutter in-hand cameras. A distal flexion mechanism absorbs differences in finger length, removing per-user adjustment.

The only quantitative result disclosed concerns hand pose tracking. Moving trackers between two mechanical stops at eight speeds from 3 to 67 cm/s, averaged over ten runs each, electromagnetic tracking rose from about 0.4 mm to 9.5 mm of mean overshoot error, while visual-inertial tracking rose from about 2.1 mm to 24.9 mm. At the highest speed that is a 60% reduction, with a narrower run-to-run spread.

Control layer

OM-1 generates robot actions directly from human motion. Because every demonstration arrives in the same format, there is no split between pre-training and post-training: a single stage trains a single policy. Each modality is processed at its sensor's native sampling rate. Beneath the policy sits a high-frequency control layer trained with reinforcement learning in simulation, handling external disturbances and system delays while motion continues during inference.

Reward AI says OM-1 picks up a new task from less than 30 minutes of human data. No success rates, public baseline comparisons or paper have been released, so the claims rest on demonstrations. Details at MarkTechPost.