Reward AI Releases OM-1, a Manipulation Policy Trained Only on Human Demonstrations

OM-1 is not available to outside developers. It is Reward AI's in-house policy, and the company has released no weights, code, dataset or API, so developers cannot run it on their own hardware yet.

Reward AI argues that human-level manipulation will not come from more teleoperated or self-collected robot data, or from more compute, because such datasets are bound to a single embodiment. The company cites Anderson's "More Is Different" in making that case. Its stated principle is "One Model, One Data Interface, Any Body," with capture, learning and control designed as one pipeline so that demonstrations recorded today can train robot bodies that do not exist yet.

The stack begins with the Omnibody Hand, a seven-degree-of-freedom wearable that extends the team's earlier DexCap work on portable motion capture. Rather than copying the human hand joint by joint, Reward AI built it around the functions it considers decisive: choosing contact points, reorienting objects in the hand, and moving between precision and power grasps. It captures thumb-index pinching, thumb and index flexion, and the coupled motion of the middle, ring and little fingers at the MCP joints. Ergonomics is treated as a data-quality issue, since a device that slips or constrains the wearer produces a compensated grasp, and a distal flexion mechanism absorbs differences in finger length so no per-user adjustment is needed.

A second component, One Data Interface, turns wearer motion into training data with no staged setup and no supervisor. Reward AI designed it with conveyor-belt sorting in mind, where a person spots, grasps and tosses an object in a fraction of a second. The glove combines high-frequency tactile sensing, proximity sensing for the pre-contact approach, and global-shutter in-hand cameras that hold context through rapid motion.

Hand pose tracking is where the company reports its first quantitative result. Visual-inertial tracking is the common default, but its accuracy at fast reversals is capped by the visual update rate, so Reward AI augmented it with electromagnetic sensing plus disturbance compensation. Moving both trackers between two mechanical stops at eight speeds from 3 to 67 centimeters per second, averaged over ten runs each, electromagnetic tracking rose from about 0.4 millimeters to 9.5 millimeters of mean overshoot error, while visual-inertial tracking rose from about 2.1 millimeters to 24.9 millimeters. That is a 60 percent reduction at the highest speed, with a narrower run-to-run spread. Force is recorded along the same trajectory, so demonstrations carry effort as well as path.

OM-1 learns to generate robot actions directly from human motion rather than routing behavior through an intermediate robot. Because every demonstration arrives in the same format, there is no split between pre-training and post-training: the first demonstration ever recorded and the newest one train a single policy in a single stage. Inputs are the glove's multimodal streams, namely images, tactile signals, inter-finger proximity and hand pose trajectories, with each modality processed at its sensor's native sampling rate rather than downsampled to a common frequency, so high-frequency tactile and motion cues survive alongside lower-frequency vision. Outputs carry motion direction, speed, force and the timing of events such as grasp initiation. Reward AI says it built a novel architecture for efficient inference, but architectural details and parameter counts are not disclosed.

Below the policy sits a high-frequency control layer trained with reinforcement learning in simulation to handle velocity- and acceleration-dependent dynamics, external disturbances and system delays. According to Reward AI, a classical controller pushed off its reference by an unexpected load never recovers, while this layer holds the reference and settles back, which is what lets a robot open a fully closed refrigerator door or lift boxes of unknown weight. The control layer runs on its own clock and continues while the policy computes the next actions, so inference latency never stalls motion, and because successive predictions may not join smoothly it optimizes the transition between them online. The same action space covers manipulation and navigation for mobile robots.

Reward AI says OM-1 picks up a brand-new task, including challenging dynamics and long horizons, from less than 30 minutes of human data, and attributes that to the integrated stack rather than to the policy alone.

Editor's Summary

Reward AI has released OM-1, an in-house manipulation policy trained only on data from humans wearing a seven-degree-of-freedom sensorized glove, with no teleoperation or on-robot data and no public weights, code, dataset or API. The company reports a 60 percent reduction in tracking overshoot error at high speed versus visual-inertial tracking, and says the policy acquires new tasks from under 30 minutes of human data. Because nothing has been released for outside use, the claims cannot yet be independently reproduced on third-party hardware.