Reward AI’s OM-1 Trains Robot Arms on Glove Data Alone, No Robot Data Needed

Reward AI, a robotics startup whose team previously built DexCap, HumanPlus, and ALOHA, has released OM-1 (Omnibody Model 1): a general-purpose robot manipulation policy trained entirely on human demonstrations captured through a sensorized wearable glove. There is no teleoperation data and no on-robot data in the training pipeline. The system is designed to run on industrial arms, legged humanoids, and wheeled mobile manipulators from a single model. As of the release, no weights, code, or API are publicly available.

What happened

Detail	Fact
System name	OM-1 (Omnibody Model 1)
Developer	Reward AI
Training data source	Human demonstrations via 7-DoF wearable glove only
Glove degrees of freedom	7 (Omnibody Hand)
New task learning time	Under 30 minutes of human data
Tracking error at 67 cm/s (electromagnetic)	9.5 mm mean overshoot
Tracking error at 67 cm/s (visual-inertial)	24.9 mm mean overshoot
Error reduction	60% at highest speed tested
Public release (weights/code/API)	None

Reward AI’s team is known for prior work on DexCap (portable motion capture), HumanPlus, and ALOHA. OM-1 follows their principle of “One Model, One Data Interface, Any Body”: a single policy trained on a single data format that can transfer to robot bodies that do not even exist yet at the time of data collection.

The core argument is that teleoperation data ties a dataset to one robot embodiment. Reward AI cites Anderson’s “More Is Different” to argue that scaling more of the same kind of data, or more compute, will not produce human-level manipulation. Instead, they built capture, learning, and control as one integrated pipeline.

How the hardware and data pipeline work

The Omnibody Hand glove

The wearable is a seven-degree-of-freedom device built around the functions that matter most in manipulation: choosing contact points, reorienting objects in-hand, and switching between precision and power grasps. It captures thumb-index pinching, thumb and index flexion, and the coupled motion of the middle, ring, and little fingers at the MCP (metacarpophalangeal) joints.

Reward AI treats ergonomics as a data-quality problem: a glove that slips or constrains the wearer produces compensated grasps that corrupt the training signal. A distal flexion mechanism handles differences in finger length automatically, so no per-user calibration is needed.

Tracking: electromagnetic beats visual-inertial at speed

Most motion-capture systems use visual-inertial tracking, but its accuracy degrades at fast reversals because it is capped by the visual update rate. Reward AI adds electromagnetic sensing with disturbance compensation on top. Across eight speeds from 3 to 67 cm/s, averaged over ten runs each, the electromagnetic tracker rose from about 0.4 mm to 9.5 mm of mean overshoot error. The visual-inertial tracker rose from about 2.1 mm to 24.9 mm over the same range, a 60% reduction at the highest speed in favour of the electromagnetic approach, with a narrower run-to-run spread.

The glove also combines high-frequency tactile sensing, proximity sensing for pre-contact approach, and global-shutter in-hand cameras. Force is recorded alongside position, so demonstrations carry effort data as well as path data.

Single-stage training, no pre-train or fine-tune split

Because every demonstration arrives in the same format regardless of when it was recorded, OM-1 trains in a single stage. There is no split between a pre-training phase and a fine-tuning phase. Inputs include images, tactile signals, inter-finger proximity, and hand pose trajectories. Each modality runs at its sensor’s native sampling rate rather than being downsampled to a shared frequency, so fast tactile and motion cues are not averaged away by slower vision data.

Outputs carry motion direction, speed, force, and timing signals such as grasp initiation. Reward AI says it built a novel architecture for efficient inference, but parameter counts and architectural details are not disclosed.

A reinforcement-learning control layer on its own clock

Below the policy sits a high-frequency control layer trained with reinforcement learning in simulation. It handles velocity- and acceleration-dependent dynamics, external disturbances, and system delays. Crucially, it runs on its own clock, so it keeps executing while the policy computes the next action. Inference latency never stalls motion. The layer also optimises transitions between successive policy predictions online, smoothing out any discontinuities. The same action space covers both manipulation and navigation, which is how OM-1 runs on wheeled mobile manipulators as well as arms.

Why it matters

The standard path to a robot manipulation policy involves collecting data with teleoperation rigs or letting robots collect their own data through trial and error. Both approaches are slow, expensive, and tied to whatever robot body the rig was built for. A system that learns purely from a human wearing a glove could, in principle, accumulate data at the pace of human labor rather than robot lab time.

The under-30-minutes claim for learning a brand-new task is the most striking number here. If it holds up across task types and environments, it represents a significant reduction compared to the hundreds or thousands of demonstrations typically required by teleoperation-based systems. The energy footprint of data collection also drops if you remove the robot from the loop entirely. For context on how energy costs accumulate as AI systems scale, our coverage of AI agent energy use versus chatbots is relevant background.

The embodiment-agnostic design is also notable. If a single dataset can train an arm, a humanoid, and a wheeled manipulator without re-collection, the economics of robot training improve considerably.

Our take

OM-1 is technically interesting, but it arrives with no benchmark numbers, no public paper, no success rates, and no external validation. Every claim in this release is demonstration-backed only, meaning Reward AI filmed robots doing things and says the model learned in under 30 minutes. That is not the same as a peer-reviewed ablation or a third-party reproduction.

The 60% tracking improvement at 67 cm/s is the one hard, quantified result in this release, and it is credible on its face given the physics of visual-update-rate limits. But whether OM-1 actually generalises to noisy real-world environments, cluttered scenes, or operators with different hand sizes remains an open question.

For businesses watching the robotics space, the direction is clear: human-demonstration pipelines are getting cheaper and faster, and the gap between human skill and robot deployment is narrowing. But OM-1 is not deployable today. No weights, no API, no licensing model. Watch for a paper or a commercial announcement before factoring this into any procurement or automation planning. If you are already mapping automation workflows for your operations, the workflow automation services we offer are a closer-term lever than waiting on robot policies.

What to do about it

1. Watch Reward AI’s publications page for a technical paper with benchmark comparisons and success rates before drawing conclusions.
2. Track whether OM-1 releases weights or an API; without that, it cannot be tested on your own hardware.
3. Note the electromagnetic-over-visual-inertial tracking result if you work with any motion-capture or wearable data pipelines; the 60% error reduction at fast reversals is replicable insight regardless of OM-1’s broader claims.
4. Review your own process bottlenecks. If repetitive manipulation tasks are a cost centre, flag this space for a revisit in 12 to 18 months when commercial robot policy products are likely to have more public validation.

The glove-only training approach is a bet worth watching, but wait for a reproducible paper before treating the demo videos as proof.

Source: Marktechpost