OM-1: a robotic policy trained only on human gestures

Reward AI presents a model learned without teleoperation or robot testing, transferable as is from an industrial arm to a humanoid, capable of assimilating a new task in less than 30 minutes of data.

Reward AI has unveiled OM-1, a robotic control model that takes the opposite view of the dominant method: instead of accumulating hours of teleoperation and testing on machines, it trains exclusively on manipulation data captured from humans wearing capture devices. No robots are needed during the learning phase.

The result is a model that transfers in zero-shot to multiple body types, from industrial arms to humanoids, and that holds up to high-contact tasks — those where the control has to deal with friction and effort, historically the most difficult to generalize. The company reports that OM-1 assimilates an entirely new task, including dynamics, with less than 30 minutes of data.

The quality of the capture determines the whole: Reward AI says it has reduced the average tracking error from 24.9 mm to 9.5 mm at high speeds by switching to electromagnetic tracking. The approach is in line with the DexCap system developed at Stanford, and aims to circumvent the bottleneck of current robotics: the cost of producing training data.