Reward AI launches OM-1 to run one policy across robot bodies

Why it matters

OM-1 targets the data reset that follows each new robot design. Reusing human demonstrations across hardware could cut deployment time and make manipulation data a durable asset.

Reward AI co-founders Zipeng Fu (@zipengfu) and Chen Wang (@chenwang_j) launched OM-1 on September 14th, pitching a robot foundation model that can control tabletop arms, industrial systems and humanoids without robot-specific fine-tuning.

The claim goes directly at one of robot learning's expensive constraints: a policy trained for one machine often needs new data and training before it can run on another. Reward AI says OM-1 instead learns a shared representation of manipulation from people wearing the company's sensor-equipped Omnibody Hand, then maps the resulting actions onto different robot bodies through a separate control layer.

Fu and Wang are building from several years of Stanford robotics research. Fu's Stanford profile lists Mobile ALOHA, Robot Parkour Learning and Rapid Motor Adaptation among his work. He studied at UCLA and Carnegie Mellon before entering Stanford's computer science PhD program, where he worked with Chelsea Finn, and previously conducted research at Google DeepMind. Wang's research profile focuses on dexterous manipulation and imitation learning under Stanford professors Li Fei-Fei and C. Karen Liu, following research positions at Google DeepMind, Nvidia and MIT CSAIL.

Their histories explain OM-1's architecture. Fu's prior work included whole-body control and Mobile ALOHA, a low-cost mobile manipulator trained through human demonstrations. Wang led DexCap, a portable motion-capture system that collected wrist and finger movement for dexterous robot policies. Reward AI describes OM-1 as a commercial-scale continuation of those two lines of research: collecting manipulation directly from people, then making the learned behavior portable across hardware.

Human data becomes the common interface

Omnibody Hand is a wearable device with seven degrees of freedom, designed around pinching, object reorientation and transitions between precision and power grasps. It records images, tactile feedback, proximity measurements and hand trajectories while a person performs a task naturally.

Those inputs feed OM-1, which outputs motion direction, speed, force and the timing of actions such as grasping. Different sensor streams remain at their native sampling frequencies, according to Reward AI, preserving fast tactile and motion signals rather than reducing every input to the frame rate of a camera.

Reward AI says the training set contains neither teleoperation data nor experience gathered by robots. Instead, the model generates robot actions directly from human motion, while a reinforcement-learning controller trained in simulation handles the dynamics, delays and physical constraints of each machine.

That division is central to the bet. Human demonstrations supply the high-level manipulation behavior. The controller absorbs embodiment-specific details at high frequency, allowing the policy to continue producing actions without pausing the robot while inference runs.

The approach also avoids dividing demonstrations into separate pretraining and post-training sets. Reward AI says every recording enters the same policy-training process and can be reused for future hardware. That would make the value of the data less dependent on whichever robot arm, gripper or humanoid happens to be available when it is collected.

The demonstrations put speed ahead of spectacle

In its launch thread, Reward AI showed OM-1 folding laundry, packaging phones with four arms, bartending, opening a refrigerator and unplugging a latched Ethernet cable. The company labels the footage as real time and says a new task can be learned from less than 30 minutes of demonstration data.

Reward AI also claims OM-1 can compensate when one arm makes a mistake, retry failed motions and stop when environmental changes make the original plan impractical. Those behaviors matter in production settings, where cycle time and recovery determine whether a robot saves labor or creates another machine that needs supervision.

The launch evidence consists primarily of Reward AI's demonstrations and technical write-up. The clips establish that the system can execute varied tasks at useful speeds. Repeatability across unseen machines, changing environments and extended production runs will determine whether the shared-policy design holds up beyond selected demonstrations.

Reward AI supplied one quantitative comparison for its data-collection hardware. In an internal test conducted at eight movement speeds and averaged across 10 runs per speed, the company says electromagnetic tracking reduced mean overshoot error at the highest speed from 24.9 millimeters to 9.5 millimeters compared with visual-inertial tracking. The result supports Reward AI's argument that fast human movement requires sensing beyond camera-based pose estimation, though it measures the capture system rather than OM-1's task success.

The broader robotics field has already moved toward cross-embodiment training through projects such as Open X-Embodiment, which standardized data collected from many robot types. Reward AI is making a narrower and potentially more economical bet: collect the behavior once from humans, then let the control stack translate it to each machine.

If that translation works reliably, robot makers can change hardware without discarding the manipulation data accumulated before the upgrade. For Fu and Wang, that is the commercial premise behind "one model, any body": the durable asset is the human behavior dataset, while robot hardware becomes replaceable infrastructure.