Reward AI’s OM-1 Policy Brings Human Hand Skills to Robots Across Arms and Humanoids

People handle boxes, doors, and tools with an ease that still frustrates most machines. A hand adjusts mid-reach when something feels heavier than expected, pulls against a sticky latch, or tosses an item into a bin without stopping to recalculate. Reward AI built OM-1, short for Omnibody Model 1, to move that kind of everyday competence into robots by recording ordinary human actions and training one policy on those recordings alone.

The company refers to OM-1 as the brain of the Omnibody stack. To accomplish this, simply wear a small, portable seven-degree-of-freedom device known as the Omnibody Hand and go about your day normally. The glove detects video from cameras on the hand, as well as pressure from touch sensors, if the space between your fingers has just been filled and the object is being touched, the precise path your hand travels, and the amount of effort you make with each movement. Then there’s electromagnetic detection on top of the visual-inertial tracking to maintain position accuracy, particularly while traveling swiftly, reducing average overshooting error by a considerable 60% at top speeds of 67 centimeters per second.

All of those different recordings are fed into one interface, and as a result, the policy can simply learn from how you move, just like a human being. There’s no need for remote control sessions or practice runs with an actual robot because the entire training procedure is simple and uncomplicated. OM1- will then instruct the robot in which direction to move, how quickly to move, how much force to apply, and when to begin or end a grab. That is then handed on to a second control layer, which employs reinforcement learning in simulation to convert it into joint commands for the specific machine in question. That layer runs on its own rapid clock, so any lag in the higher decisions never pauses the motion, and it smooths the handoff between successive predictions to keep actions continuous during quick tosses or swings


The same policy is utilized on industrial arms, humanoids, and mobile platforms, allowing for the development of a single policy that works across all machines. Then they demonstrate the robots performing a variety of tasks, such as sorting items off a conveyor belt as it moves, packing phones with two arms working together, folding laundry, unscrewing fittings with one arm while another steadies the part, unplugging an ethernet cable, making a drink, and even opening a fully closed fridge door, all on their own. The videos are all in regular speed and tagged as fully autonomous operation. Apparently, the goal is to teach a new activity in less than 30 minutes, even if it requires lengthier sequences and you’re not sure what forces you’ll be working with.


Because every demonstration arrives in the same format, training stays in a single stage from the first recording onward. Adding more human data improves performance for robots already in use and for platforms that have not been built yet. Reward AI has not released model weights, code, a technical paper, or measured success rates on any of the shown tasks, so outside checks of the full results remain limited. The only published number compares the two tracking methods on the glove itself. The method still removes the usual requirement to collect fresh practice data on every new robot body before it can attempt the work.