The human hand is one of nature’s great pieces of engineering, capable of cracking an egg, tossing a ball, folding a shirt or pulling open a stubborn refrigerator door, often without its owner giving any of those movements much thought. For robots, the simple business of reaching, touching, gripping and adjusting on the fly remains a difficult act to follow.

Reward AI, a robotics company based in the San Francisco Bay Area, thinks the best way to close that gap may be to stop teaching robots how to manipulate objects like robots and start teaching them how to do it like people.

Reward AI has developed OM-1, its first general-purpose robot foundation model, as part of a broader system called Omnibody. The technology is built around the idea of “One Model, One Data Interface, Any Body.”

Zipeng Fu, the co-founder and CEO of Reward AI, has a doctorate in computer science from Stanford University, where he conducted research in the Stanford AI Lab, and worked as a researcher at Google DeepMind. In 2024, when he was a graduate student, he was among a team at Stanford that created a low-cost mobile robot that quickly learned how to sauté shrimp, put away dishes and clean up spills.

“When Chen and I founded Reward AI, we wanted to help bring robots into everyday use,” he said, referring to co-founder and chief technology officer Chen Wang, also a Stanford AI Lab graduate. “That goal runs through our whole team. We focus on two challenges central to deploying robots at scale: intelligence and performance. A robot earns its place in a factory, a kitchen, or a warehouse through what it does every day, in any body, at full speed, alongside others.”

Instead of collecting specialized training data for every robot, Reward AI captures natural human manipulation and teaches a single AI model to translate those movements into actions across different machines, including tabletop and industrial robotic arms and humanoids.

Building on DexCap, earlier Stanford University research for capturing dexterous human manipulation, Reward AI created Omnibody Hand, a wearable device designed to record what people actually do with their hands. Rather than forcing a person to operate a robot remotely, the system allows someone to work naturally while capturing the physical intelligence embedded in movements such as pinching, twisting, sliding, grasping and repositioning an object.

Its seven-degree-of-freedom design does not attempt to mechanically duplicate every joint in the human hand. Instead, Reward AI concentrated on functions that matter for manipulation, including thumb-and-index-finger precision, moving objects within the hand and shifting smoothly between delicate pinches and powerful grasps.

The system also records much more than movement. Cameras provide visual information, while proximity and tactile sensors capture what happens as the hand approaches and touches an object. Electromagnetic sensing helps track rapid movement, and force measurements record how much effort a person uses.

That data feeds OM-1.

The model learns directly from human demonstrations, according to Reward AI, without teleoperation or robot-generated training data. Images, touch, proximity, hand position, force and movement over time give OM-1 a multidimensional picture of an interaction. The model then produces commands governing such things as a robot’s direction, speed, force and the timing of a grasp.

Consider unplugging an Ethernet cable. A robot must recognize the connector, position its fingers, locate and depress a small locking tab, maintain the right pressure and pull at precisely the right moment. Reward AI says OM-1 can reliably perform that kind of task.

The company has demonstrated the system performing other activities, including opening refrigerator doors, folding laundry, packaging phones and handling objects of different weights. Reward AI says OM-1 can also learn a new task involving complex dynamics and multiple steps from less than 30 minutes of human demonstration data.

Many robotic AI systems are closely tied to the machine on which they were trained. Reward AI says OM-1 can instead transfer what it has learned across different robotic bodies without robot-specific fine-tuning.

That could become important as robotics hardware evolves. A human demonstration recorded today, the company argues, could eventually help train machines that have not yet been designed.

The company is also exploring what happens when several robots share the same model. Reward AI has demonstrated coordinated manipulation involving multiple machines and envisions robots eventually dividing tasks, handing objects to one another and adapting when another robot makes a mistake.

“Much of the most valuable physical work is bigger than any single body: carrying furniture, assembling large structures, running a kitchen at rush hour,” Fu said. “Humans solve this with teams, and teams run on mutual understanding. We are building that understanding into robots, beginning with quadmanual manipulation, where four robots coordinate and collaborate, and extending to mixed teams that read each other’s intent, hand off objects, divide roles, and recover together when something slips, all without scripted choreography. Because our robots share one brain, collaboration is not a communication protocol bolted on afterward; it is closer to two hands of the same body. And this is where robots go where humans cannot. People can’t share minds. Robots can. What one learns, every robot knows.”