Reward AI Unveils OM-1: A Robot Policy Trained Exclusively on Human Demonstrations

Industry Brief is building the Omnibody stack around a single principle: One Model, One Data Interface, Any Body. The News Reward AI, a robotics startup, has announced the release of OM-1 (Omnibody Model 1), a general-purpose manipulation policy that learns directly from human demonstrations captured using a sensorized glove. A significant departure from conventional methods, OM-1 requires no teleoperation data or on-robot data for its training. The policy is designed to operate on diverse robot platforms, including industrial arms and humanoids, at human speeds. What It Is OM-1 is an in-house developed general-purpose manipulation policy by Reward AI. It functions through an integrated system that includes the 'Omnibody Hand,' a 7-degree-of-freedom (DoF) wearable sensorized glove that captures human motion and contact points. This 'One Data Interface' approach converts human actions into training data without the need for staged setups or supervisors. While the technology has been released, Reward AI has not yet made the weights, code, dataset, or an API publicly available. Key Highlights Learns exclusively from human demonstrations captured via a 7-DoF sensorized glove, the 'Omnibody Hand.' Training explicitly avoids teleoperation or on-robot data, reducing reliance on robot-specific data collection. Aims to be a general-purpose manipulation policy applicable to various robot embodiments, including industrial arms and humanoids. Reportedly capable of learning new, complex…