Reward AI unveils om-1: a robot policy trained solely on human demonstrations

Reward AI, a robotics startup known for prior projects like DexCap and ALOHA, has launched OM-1, a general-purpose manipulation policy trained exclusively on human demonstrations. OM-1 operates on industrial arms and humanoids at human speed, without requiring teleoperation or on-robot data. The system is built on the principle of 'One Model, One Data Interface, Any Body.' OM-1's training data comes from a wearable called the Omnibody Hand, which captures motion with a seven-degree-of-freedom design, focusing on key functions like contact points and object reorientation. The device adapts to different users without calibration, using a distal flexion mechanism. Reward AI claims that demonstrations recorded today can train robots that do not yet exist, emphasizing an integrated pipeline for capture, learning, and control. The policy learns directly from human motion, with inputs from multimodal sensors and outputs including motion direction, speed, and force. A high-frequency control layer, trained via reinforcement learning, handles dynamic challenges like opening a closed refrigerator door. Despite the lack of public benchmarks, Reward AI asserts that OM-1 can learn new tasks from less than 30 minutes of human data.