- Published on
A Robot Policy Trained on Human Demonstrations Only, With No Teleoperation or On-Robot Data
- Authors
- Name
- DeepHorus
Reward AI, a robotics startup founded by researchers behind DexCap, HumanPlus, and ALOHA, has unveiled OM-1 (Omnibody Model 1), a general-purpose manipulation policy trained exclusively on human demonstrations captured with a sensorized glove. The system runs on industrial arms and humanoids at human speed without any teleoperation or on-robot data in its training pipeline. The company describes its approach as "One Model, One Data Interface, Any Body."
The startup argues that scaling teleoperated or self-collected robot data will not lead to human-level manipulation, citing Anderson's "More Is Different." Instead, Reward AI integrates capture, learning, and control into a single pipeline, allowing demonstrations recorded today to train robot bodies that do not yet exist. OM-1 is not publicly deployable; no weights, code, dataset, or API have been released.
The hardware foundation is the Omnibody Hand, a seven-degree-of-freedom wearable that extends the team's earlier DexCap work. Rather than mimicking every human joint, it focuses on contact-point selection, in-hand reorientation, and transitions between precision and power grasps. A distal flexion mechanism accommodates different finger lengths, eliminating per-user calibration.
Reward AI's first quantitative result concerns hand-pose tracking. In tests moving two trackers between mechanical stops at speeds from 3 to 67 cm/s, electromagnetic tracking reduced mean overshoot error from 24.9 mm to 9.5 mm at the highest speed compared with visual-inertial tracking—a 60 percent improvement. Force is recorded along the same trajectory, so demonstrations include effort as well as path.
The One Data Interface captures contact at human speed, targeting tasks like conveyor-belt sorting, where a person grasps and tosses an object in a fraction of a second. It combines high-frequency tactile sensing, proximity sensing, and global-shutter in-hand cameras. This design requires no staged setup or supervisor.
OM-1 learns to generate robot actions directly from human demonstrations in a single-stage training process. While the policy is not yet available to developers, Reward AI positions it as evidence that human-level manipulation can emerge from human data alone. The company's future roadmap suggests that embodiment-agnostic learning could reshape how robot policies are trained, though commercial release remains uncertain.
