Reward AI Releases OM-1 Robot Foundation Model Learning from Human Hand Data
Reward AI released the OM-1 robot foundation model, which learns exclusively from human hand operation data without requiring teleoperation or robot-specific data. The model can be deployed directly to various robot types including desktop arms, industrial arms, and humanoid robots after training. A new task requires under 30 minutes of human demonstration data to learn.
Key insights
Novel Training Approach Without Teleoperation
OM-1 differs from mainstream robot foundation models like Physical Intelligence's π0 and Figure AI's Helix by avoiding teleoperation data. Instead, it uses the Omnibody Hand wearable device, a 7-DOF system that captures multimodal data (visual, tactile, force) during human daily activities like cooking and organizing. This approach bypasses the high cost and hardware-specific limitations of teleoperation-based data collection.
Demonstrated Capabilities and Spontaneous Behaviors
Demo videos show OM-1 completing tasks in real time without acceleration, including phone packaging (four-arm collaboration), bartending, folding clothes, and unplugging network cables requiring precise button pressing. The model exhibits spontaneous behaviors such as automatic compensation between arms during multi-arm tasks, retrying after external interference, and stopping when environmental changes become too significant.
Development Stage and Commercialization Status
OM-1 is currently in the research and development demonstration phase. Reward AI has not announced a commercialization timeline or pricing. The company originated from Stanford University's DexCap project, a portable hand motion capture system that records fine finger movements during daily operations.