Reward AIThe team behind a new robotics company has been named prior work Released include DexCap HumanPlus and ALOHA OM-1, short for Omnibody Model 1. OM-1 has a general-purpose manipulating policy which learns by humans using a sensorized gloves, and then is run on industrial arms or humanoids with human speed. It is important to note that neither teleoperation nor robot-related data are included in the training. The system follows one principle, ‘One Model, One Data Interface, Any Body,’.
Is it deployable? Reward AI has its own policy, OM-1. Developers cannot use it yet because no weights, data, API, code or dataset have been made available.
Why skip robot data?
Robot foundation policies are based on data collected by teleoperated robots or robots that self-collect their own. This means the dataset is tied to a particular embodiment. Reward AI asserts that the only way to achieve human-level control is through more computing or data. “More Is Different.” As a result, demonstrations that are recorded today will be able to train robots which do not yet exist.
Omnibody Hand: A 7-DoF Wearable
Start with the stack Omnibody HandThe team has developed a wearable device that extends their earlier DexCap Work on portable motion-capture. Rather than copying the human hand joint by joint, it is a seven-degree-of-freedom design built around the functions that matter: choosing contact points, reorienting objects in-hand, and moving between precision and power grasps. The MCP joint captures the thumb-index pinching motion, the thumb and index flexion and the coupled movement of the little, middle and ring fingers.
Ergonomics can be viewed as an issue of data quality: any device which slips on the wearer or restricts them will result in a compensated grasp. Distal flexion mechanisms compensate for differences in the length of fingers, meaning that no user-specific adjustment is required.
One Data Interface Wearer movement is transformed into training data without a staged set-up or supervisor. The conveyor belt sorting target was chosen because a person can spot, pick up, and then toss an item in just a fraction of second. For the entire interaction, the glove integrates high-frequency tactile sensors, proximity sensor for the precontact approach and in-hand global-shutter cameras to hold context even during rapid motion.
Reward AI’s first quantifiable result is the hand pose tracking. Visual-inertial is often the default tracking method, but accuracy for fast reversals depends on the rate of visual updates. It is enhanced by Reward AI with electromagnetic sensing and disturbance compensation. The average overshoot of both trackers was reduced by 60% at high speeds, with the visual-inertial tracking averaging a reduction from 2.1 to 24.9mm. Since force and path are recorded on the same trajectory for demonstrations, they can be compared.
OM-1 – One policy, single-stage training
OM-1 The robot can be taught to perform actions directly from human movements, without having to route the behavior through an intermediary robot. Every demonstration is the same, so there’s no need to separate pre-training from post-training. Both the oldest demonstration and the most recent one teach the policy simultaneously.
The glove receives inputs in the form of multimodal streams, including images, tactile signals and inter-finger distances. Also, hand position trajectories are also included. Inputs are the glove’s multiple-modal streams: images, tactile cues, proximity between fingers and hand poses. The outputs include motion direction, velocity, force and timing for events like grasp initiation. Reward AI claims to have built a novel inference architecture, but architectural details and parameter count aren’t disclosed.
The RL layer on its own clock: Control any body
Below the policy sits a high-frequency control layer trained with reinforcement learning in simulation to handle velocity- and acceleration-dependent dynamics, external disturbances, and system delays. The classical controller is never able to recover from an unexpected load. However, the high-frequency control layer can hold its reference, and then settle back.
As the control layer is on its clock and continues to run while the policy calculates the next action, inference latencies never slow down motion. The transition from one prediction to the next is optimized online because successive predictions might not merge smoothly. Mobile robots can also manipulate and navigate using the same space.
The results of the study are:
Reward AI states that OM-1 can pick up a completely new task from just 30 minutes worth of data. This includes challenging dynamics, long horizons and complex tasks. Its about page It is stated that all clips published run at 1x speed, that the model encompasses humanoids with legs and arms and wheels, as well as arm-based manipulators. The claims made are not backed by benchmarks but rather demonstrations.

