Reward AI’s OM-1 cuts robot training data needs by skipping teleoperation

Reward AI just dropped OM-1, a general-purpose manipulation policy that bypasses the industry’s heavy reliance on teleoperated robot data. The startup, whose team built DexCap and HumanPlus, argues that binding datasets to specific robot bodies is a dead end. Instead, they propose a pipeline where humans wear a sensorized glove, perform tasks at natural speed, and the model learns directly from that human motion. No robot data. No staged setups. Just human demonstration translated into robotic action.

The hardware backbone is the Omnibody Hand, a seven-degree-of-freedom wearable designed around function rather than anatomical mimicry. It captures pinching, reorienting, and grasping without needing per-user adjustments. The real technical leap here is in the tracking accuracy during rapid motion. Reward AI reports that their electromagnetic sensing system reduced mean overshoot error to 9.5 mm at high speeds, compared to 24.9 mm for standard visual-inertial tracking. That 60% reduction matters because it means the data remains clean even when workers move at conveyor-belt pace.

OM-1 processes these multimodal streams—tactile, proximity, vision, and pose—at their native sampling rates. This preserves high-frequency cues that usually get lost in downsampling. The policy then feeds into a separate reinforcement learning control layer that handles dynamics and disturbances in simulation. This decoupling allows the robot to recover from unexpected loads, like lifting a box heavier than anticipated, without stalling. The system claims to master new tasks with less than 30 minutes of human data.

But there is a catch for developers: you can’t use it yet. OM-1 remains an in-house policy with no released weights, code, or API. While the principle of ‘One Model, One Data Interface, Any Body’ sounds promising for future humanoid deployment, the current lack of accessibility keeps it firmly in the research realm. It proves the concept works, but until Reward AI opens the stack, this remains a impressive demo rather than a deployable tool for the broader robotics community.