Reward AI's OM-1 pushes each robot's dynamics into a simulation-trained controller

OM-1's output stops short of joint commands. It takes hand trajectories, images, tactile feedback and proximity readings from the wearable ^[4] and outputs motion direction, speed, force and the timing of actions such as grasping ^[5]. A reinforcement-learning controller trained in simulation turns that into motion a particular machine can execute, and it carries that machine's dynamics, delays and physical constraints ^[8]. Reward AI says the controller absorbs those embodiment-specific details at high frequency, so the robot keeps moving while the policy's inference is still running ^[9].

One line in the technical write-up matters more than the portability claim. Reward AI says each sensor stream stays at its native sampling frequency, with no down-sampling to a camera's frame rate, keeping the fast tactile and motion signals intact ^[6]. The wearable is built around pinching, object reorientation and the transition between precision and power grasps ^[3]. Those are contact events, and resampling them to video rate loses the transient the grasp decision depends on.

The 30-minute figure is Reward AI's measurement of Reward AI's tasks, on its own wearable and its own controllers ^[12]. For it to hold on someone else's line, the task has to be performable by a human hand wearing a seven-degree-of-freedom device ^[3]. And a simulated model of the target machine has to be good enough to train a controller against ^[8]. The recovery behaviour has to survive the same transfer. Compensating when one arm makes a mistake, retrying a failed motion and stopping when the environment changes appear in the launch thread as claims, without measured rates ^[13]. Reward AI did not say how long controller training takes for a machine it has not used before.

The archive claim matters commercially. Reward AI says it does not divide demonstrations into separate pretraining and post-training sets, so every recording enters the same policy-training process and stays reusable for hardware acquired later ^[10]. Since the training set holds no teleoperation and no robot-collected experience ^[7], the only per-robot artifact in the pipeline is the simulation-trained controller ^[19]. That ties the value of the archive to the Omnibody Hand's sensing format.

Both founders have been at this specific problem for years. Chen Wang led DexCap, a portable motion-capture system that collected wrist and finger movement for dexterous robot policies ^[16]. Zipeng Fu's work includes Mobile ALOHA, a low-cost mobile manipulator trained through human demonstrations ^[15]. Reward AI describes OM-1 as a commercial-scale continuation of those two lines: collect manipulation from people, then make the behaviour portable across hardware ^[17][18].

runtimewire.com cites Reward AI's post on X as its primary source ^[20]. It writes that the clips establish the system can execute varied tasks at useful speeds, and that repeatability across unseen machines, changing environments and extended production runs will determine whether the shared-policy design holds ^[21]. Of the clips in the thread, unplugging a latched Ethernet cable is the most legible: the latch either releases or it does not ^[11].