Reward AI's OM-1 learns robot skills from human glove data alone, with no teleoperation or robot data

Reward AI released OM-1 on September 14, 2026, a general-purpose robot policy trained only on human demonstrations recorded with a 7-degree-of-freedom wearable called the Omnibody Hand, with no teleoperation data and no on-robot data. The company, co-founded by former Stanford researchers Zipeng Fu and Chen Wang, says OM-1 learns a new task from less than 30 minutes of human data and runs as one policy on industrial arms, humanoids and wheeled mobile manipulators.

Why it matters: most robot foundation models still depend on teleoperation, where an operator drives the robot to generate training data. A policy that trains on people wearing a glove would remove the robot from the data-collection loop entirely, if the claims hold up outside the company's own demos.

What is OM-1 and how is it trained?

Reward AI describes OM-1 as "a general-purpose robot policy learned directly from human data, without teleoperation or on-robot data," and summarises its approach as "One Model, One Data Interface, Any Body." According to MarkTechPost's review of the technical post, the model takes images, tactile signals, inter-finger proximity and hand-pose trajectories, each at its native sensor sampling rate. RuntimeWire reports that it outputs motion direction, speed, force and the timing of actions such as grasping, without robot-specific fine-tuning. A separate reinforcement-learning control layer handles dynamics, disturbances and latency compensation on each robot.

The company's headline claim is data efficiency: OM-1 "picks up a brand-new task, including challenging dynamics and long horizons, from less than 30 minutes of data." It also says the policy operates at human speed. Both claims are company-reported.

What is the Omnibody Hand?

The Omnibody Hand is a sensorized glove with seven degrees of freedom, which MarkTechPost describes as an extension of DexCap, the portable motion-capture system Wang led at Stanford. Its design groups the finger joints to capture thumb-index pinching, finger flexion and power grasps. Reward AI reports that its electromagnetic hand-pose tracking cut mean overshoot error to 9.5 mm at 67 cm/s, against 24.9 mm for a visual-inertial baseline, a 60% reduction at the highest speed tested.

Who founded Reward AI?

Zipeng Fu, the co-founder and CEO, completed a Stanford PhD advised by Chelsea Finn and previously worked at Google DeepMind. He co-led Mobile ALOHA, the low-cost bimanual mobile manipulator, and HumanPlus, a humanoid system that learns by shadowing humans. Chen Wang studied at Stanford under Fei-Fei Li and C. Karen Liu, held positions at Google DeepMind, Nvidia and MIT CSAIL, and led DexCap.

Reward AI has not disclosed funding, investors or headcount. MarkTechPost notes that no weights, code, dataset or API have been released, so the cross-embodiment and 30-minute claims cannot yet be checked independently.

The bet runs against the current spending pattern in robot data, from teleoperation interfaces such as the one Enigma raised $71M to build to the simulation route covered in our sim-to-real analysis.