Robotics startup Reward AI has unveiled OM-1, a foundation model the company says can control almost any robot body without ever being trained on a robot. Instead of teleoperation rigs or robot-specific datasets, OM-1 learns entirely from people performing everyday tasks while wearing the company's Omnibody Hand, a seven-degree-of-freedom sensorized glove. That human data alone is enough for the model to generalize zero-shot across tabletop arms, industrial arms, and full humanoids. Reward AI showcased the system handling long, complex sequences, including packaging phones with four hands working at once, mixing drinks behind a bar, folding laundry, and even unplugging a locked Ethernet cable, all executed in real time at a pace the company says approaches human speed and fluency.

The Omnibody Hand itself was built less to mimic human anatomy and more to capture what actually matters for manipulation, prioritizing useful contact points and smooth transitions between precision and power grips, so the system stays relevant even as robot hardware keeps changing. OM-1 also supports multiple robots working together, and Reward AI reports emergent behaviors during these sessions, like one unit automatically compensating when another makes a mistake. The team behind it has deep roots in the field, having previously built DexCap, HumanPlus, and ALOHA, well-known projects in human-to-robot learning research. For now, though, OM-1 remains Reward AI's in-house system, no public weights, code, or API have been released, so outside developers cannot yet test it on their own machines.