Published 15 September 2026
Reward AI published OM-1 on 14 September 2026, a general-purpose robot manipulation policy the company says picks up a new task from under 30 minutes of human demonstration. What makes OM-1 unusual is the training data: no teleoperation and no robot experience of any kind, only recordings of people wearing a sensor glove. No weights, code or paper have been released, and Reward AI has published no success rate for any task.
How does OM-1 learn without robot data?
Nearly every robot policy today is trained on teleoperation. A person drives the robot with a controller or a motion-capture rig, and the model learns from what the robot did. That data is slow and expensive to collect, and it mostly teaches you about the one robot you collected it on.
Reward AI’s answer is a wearable it calls the Omnibody Hand, a seven-degree-of-freedom device rather than a copy of every finger joint. A person wears it and does the task at normal speed. It records camera images, touch pressure, the distance between fingers just before contact, and the path the hand takes. OM-1 learns to produce robot actions directly from that, with no robot involved in training.
Underneath the policy sits a separate control layer, trained in simulation with reinforcement learning, that turns OM-1’s intended motions into joint commands for whatever machine it is on. That layer runs on its own clock, so when the policy takes longer to think, the robot does not stutter. Reward AI says the same setup covers industrial arms, legged humanoids and wheeled mobile manipulators.
What Reward AI actually measured
Only one experiment in the release produces numbers, and it tests the glove rather than the model. Reward AI mounted its electromagnetic tracker and a standard visual-inertial tracker on the same rig and moved them between two fixed stops at eight speeds, ten runs each. At the fastest speed, 67 centimetres per second, the electromagnetic tracker overshot by 9.5 mm on average against 24.9 mm for visual-inertial, a 60% reduction. At the slowest speed both stayed under 2.1 mm.
That is a real result and a sensible piece of engineering. It is also a sensor comparison. It tells you the glove captures fast motion cleanly, which matters because the whole approach depends on the recording being accurate. It tells you nothing about whether a robot running OM-1 completes a task.
How does OM-1 compare to GR00T N1.7 and Gemini Robotics 2?
The gap between OM-1 and its closest rivals is not really in the claims. It is in what each company was willing to publish.
| What the release includes | Reward AI OM-1 | The published alternative |
|---|---|---|
| Training data volume | not stated | 20,854 hours of human video, NVIDIA Isaac GR00T N1.7 |
| Parameter count | not stated | 3 billion, GR00T N1.7 |
| Task success rates | none | 92% unscrewing a bulb, 44% tying a bin bag, Gemini Robotics 2 |
| Weights and code | not released | Apache 2.0 on Hugging Face, GR00T N1.7 |
| Technical paper | none | arXiv:2604.15483, Physical Intelligence π0.7 |
| Published measurement | sensor tracking error | robot task completion rates |
The figures in the right-hand column are each company’s own, so they carry the same self-reported caveat as Reward AI’s. The difference is that they exist. GR00T N1.7 is the closest comparison, because NVIDIA is making the same bet on human video instead of teleoperation, and it published a scaling curve showing that going from 1,000 to 20,000 hours more than doubles average task completion. Reward AI has not said how many hours it used.
What this means
If you buy robots, ignore this release for now. There is nothing to purchase, nothing to download and no number showing a robot finishing a job. The single measurement Reward AI published compares two motion sensors, which belongs in a hardware note rather than in the announcement of a foundation model. The team also has no published funding, location or headcount, and the founders appear on the site by first name only.
If you build robot policies, watch this closely. The no-teleoperation claim is the one thing here worth arguing about, and the team behind it wrote DexCap, HumanPlus and ALOHA, which is about as strong a track record as this field offers. Learning a task from half an hour of a person simply doing that task would change the economics of robot data collection, which is the binding constraint on the whole field right now. The test is simple and has not been run in public: pick one task, state a success rate over a stated number of trials, and say which robot it ran on. Until that number appears, OM-1 is a demonstration reel with a good pedigree.
For more information, visit the official announcement of OM-1 on the Reward AI blog.
