OM-1: robot foundation model trained purely on human data

On September 15, Reward AI officially launched the Omnibody tech stack and its first robotics foundation model, OM-1. Founder Zhu Yifeng (Zipeng Fu / @yifengzhuut) explained that OM-1's core principle is "One Model, One Data Interface, Any Body": a single model can zero-shot control desktop robotic arms, industrial arms, and even humanoid robots (paired with 7-DoF Omnibody Hands) without any fine-tuning for specific robots.

Confirmed

• Training: learns directly from human manipulation data, requiring no teleoperation data or robot embodiment data
• Capabilities: near-human-level dexterity and efficiency, with support for multi-robot collaboration; all official demo videos are played in real time with no speedup
• Emergent behaviors: dual arms learn to compensate for each other's mistakes during tasks; the model can decide when to retry, when to adapt to adversarial disturbances, and when to stop when the environment changes too drastically
• Task efficiency: long-horizon tasks are compressed into short ones—four-handed phone packaging, cocktail mixing, and laundry folding complete in under 30 seconds; the team stresses that "efficiency is ROI, and ROI determines whether robots can be deployed"
• The release includes both the Omnibody tech stack and the OM-1 model, covering a unified data interface spanning robotic arms to humanoids

Why It Matters

• With hardware iterating month by month, zero-shot cross-embodiment generalization, if validated, could dramatically cut the cost of retraining/fine-tuning for every embodiment in robot deployment
• Using human manipulation data rather than teleoperation data, combined with emergent behaviors like dual-arm mutual error correction, points toward a path closer to a general-purpose robot assistant
• The "efficiency is ROI" narrative ties research capability directly to commercial deployment—a key argument in the robotics foundation model race

2026-09-15 ~ 2026-09-15 · 14 related posts

• [source] Reward AI debuts OM-1 robot foundation model, zero-shot generalizing to any robot — yifengzhu_ut · 2026-09-15
• [source] RewardAI's OM-1 zero-shot generalizes across arms and humanoids — zipengfu · 2026-09-15
• [source] Reward AI's OM-1 does quadmanual packaging, bartending, laundry folding in under 30 seconds, real-time — zipengfu · 2026-09-15
• Reward AI launches OM-1: one robot policy zero-shot across arms, industrial arms and humanoids — zipengfu · 2026-09-15
• Reward AI's OM-1 shows emergent behaviors: arms compensating for each other's mistakes — zipengfu · 2026-09-15
• OM-1 robot foundation model generalizes zero-shot across arms and humanoids from human data — zipengfu · 2026-09-15
• RewardAI unveils OM-1 robot foundation model trained on human data only, zero-shot across robots — zipengfu · 2026-09-15
• Physical Intelligence unveils OM-1, a robot foundation model trained purely on human data — zipengfu · 2026-09-15
• Physical Intelligence launches OM-1 robot foundation model with zero-shot generalization — zipengfu · 2026-09-15

5 near-duplicate retellings: zipengfu · zipengfu · zipengfu · zipengfu · zipengfu