RewardAI unveils OM-1 robot foundation model trained on human data only, zero-shot across robots
zipengfu · x · 2026-09-15
RewardAI (with researchers Chen Wang, Yifeng Zhu, Zipeng Fu et al.) introduced OM-1, its first robot foundation model that zero-shot generalizes to tabletop arms, industrial arms, and humanoids. Key points:
- Trained on human manipulation data only — no teleoperation, no on-robot experience, and no pre-train/post-train split: the first-ever demo and the newest one run the same single-stage-trained model
- Full-stack design around one constraint — matching human speed: Omnibody Hand has 7 DoF chosen for function (contact-point selection, in-hand reorientation, precision vs. power grasps) rather than joint-by-joint human mimicry
- One data interface: in-hand cameras combine tactile + proximity + global-shutter sensing; EM-augmented pose tracking cuts high-speed overshoot 60% vs. visual-inertial, keeping data usable at full human pace
- Control: RL-trained in sim, running at high frequency on its own clock so inference latency never stalls motion; supports multi-robot collaboration
The team claims near human-level dexterity and efficiency.
Related event: OM-1: robot foundation model trained purely on human data(14 posts)→
More from Embodied
- RewardAI unveils OM-1 robot foundation model with zero-shot generalization across robots — zipengfu · 2026-09-15
- Unitree's new dual-axis robot neck mocked: tilt-then-pan kinematics force absurdly long neck — carlosdponx · 2026-09-15
- Reward AI debuts Omnibody Hand & OMI-1, pushing hardware-algorithm co-design for cross-embodiment dexterity — zipengfu · 2026-09-15
- Robotics commentator Herbert joins RoboStrategy to host new robotics podcast — Rewkang · 2026-09-15
- AgentSTAR tracks and reconstructs articulated objects from monocular video through occlusion — notmahi · 2026-09-15
- Valve launches Steam Frame wireless VR headset in 256GB and 1TB models — Scobleizer · 2026-09-15