Reward AI just launched its new model, OM-1! The core team comes from Stanford Robotics/Embodied AI, and the technology directly builds on Stanford's DexCap project. The team introduced OM-1, a robotic strategy entirely learned from humans. No teleoperation (remote control data) or robot-specific data is required. One model works for any robot (cross-robot compatibility). From the videos, you can see impressive speed and flexibility (compared to many previous robot demo videos). **Data Collection:** A wearable 7-DOF robotic hand (Omnibody Hand) allows people to naturally demonstrate actions while working, cooking, or sorting. Sensors record tactile, distance, force, and hand posture data at human speed. **Learning:** OM-1 is trained solely on this human data. No teleoperation or robot-specific experience is needed. It can then run on various robots, from industrial robotic arms to humanoid robots, and scaling only requires adding more human data. **Control:** A high-frequency control layer trained in simulation using reinforcement learning (RL) enables these actions to be executed on any robot. OM-1 can master a brand-new task with less than 30 minutes of data. **Team:** - **Zipeng Fu (付子鹏)** Cofounder & CEO Background: Stanford PhD, Google DeepMind, CMU - **Chen Wang (王晨)** Cofounder & CTO Background: Stanford PhD; previously at Google DeepMind, NVIDIA, MIT CSAIL Lead author of DexCap (wearable hand motion capture for training dexterous manipulation from human operation data) - **Yifeng Zhu** Member of Technical Staff Background: UT Austin PhD Previous work on Stanford DexCap: https://dex-cap.(github.io)/