Reward AI released OM-1 ( Omnibody Model 1) —, a universal manipulation policy trained on human demonstrations using a wearable glove with seven degrees of freedom. As reported by MarkTechPost, teleoperation and data collected on robots were not used during training.
According to the source, the policy operates on industrial manipulators and humanoid robots at human speed, mastering a new task in less than 30 minutes of data. The system combines electromagnetic hand tracking with a reinforcement learning-based control layer; for tracking, it claims 60% less trajectory overshoot compared to visual-inertial methods at a speed of 67.
The practical implication of this approach is the ability to transfer human demonstrations to various robotic platforms without teleoperation. This is an analytical inference rather than a confirmed deployment result: weights, code, and the API for OM-1 are not yet publicly available.
editorial commentary
Why it matters
A likely consequence is increased interest in training robots on human demonstrations instead of telepresence. The next observable signal will be the publication of weights, code, an API, or independent tests of OM-1. Significant uncertainty remains due to the single source and lack of primary materials.