Reward AI

Reward AI has released OM-1, short for Omnibody Model 1, a general-purpose manipulation policy that learns from humans wearing a sensorized glove and then runs on industrial arms and humanoids at human speed. The launch marks the company's exit from stealth and targets robotics teams that have been bottlenecked by the cost of collecting teleoperated demonstrations on every new hardware platform. The startup, spun out of Stanford's DexCap project, says OM-1 runs zero-shot across industrial arms and humanoids and can learn long-horizon tasks with under 30 minutes of human demonstration data. The data pipeline centers on the Omnibody Hand, a wearable device people use while doing normal work. A wearable 7-DoF hand lets people demonstrate manipulation naturally while working, cooking, or sorting, with sensors recording tactile, proximity, force and hand pose at full human speed. That single stream of human data is what powers the "one model, any body" pitch behind the tweet, with the same policy driving tabletop arms, industrial arms, and humanoids without robot-specific retraining, plus multi-robot collaboration demos. The team's prior work includes DexCap, HumanPlus, and ALOHA , three of the more widely cited human-to-robot and low-cost manipulation projects of the past few years. Reward AI is based in the SF Bay Area, and cofounder and CEO Zipeng Fu was advised by Chelsea Finn at Stanford and previously worked at Google DeepMind. For investors and operators tracking the shift from teleop-heavy data collection to human-worn capture, this launch is a concrete bet that a foundation model trained only on human hands can be the shared brain for a fragmented robot hardware market.