Reward AI's OM-1 learns manipulation from glove recordings alone

single source· 1 articles · confidence: low · first seen 2026-09-14 21:08 UTC

What this means for you

Nothing to act on yet: no weights, no code, no API. If the approach holds, the thing worth watching is where the training data comes from — a glove rather than the robot — because collecting demonstrations on real hardware is the costly part. There is no published evaluation to check any of it against.

Reward AI has released OM-1 (Omnibody Model 1), a manipulation policy — the model that decides how a robot moves — trained only on recordings from a wearable glove with seven tracked joint angles, with no teleoperation and no data from the robot itself. The company says it runs on industrial arms and humanoids at the speed a person would move and learns a new task from under 30 minutes of data; it pairs electromagnetic hand tracking with a control layer trained by reinforcement learning that runs on its own clock. No weights, code or API are public, and no third-party evaluation has been released.

Key facts

  • ·Reward AI released OM-1 (Omnibody Model 1), a general-purpose manipulation policy trained entirely on human demonstrations captured with a 7-DoF wearable glove source
  • ·No teleoperation data and no on-robot data were used to train the policy source
  • ·Reward AI says a new task can be learned from under 30 minutes of data source
  • ·The policy runs on industrial arms and humanoids at human speed source
  • ·Electromagnetic hand tracking is reported at 60% lower overshoot than visual-inertial tracking at 67 cm/s, with no test conditions or evaluation date given source
  • ·No weights, code or API are public as of the 14 September 2026 announcement source

What the sources say

  • MarkTechPost — Announcement post describing a glove-trained manipulation model and noting that nothing has been released.

Sources

The original reporting. Follow these — they did the work.

← the wire