OM-1 (Omnibody Model 1)

OM-1, short for Omnibody Model 1, is a general-purpose robot manipulation policy announced by Reward AI on 14 September 2026 ^[1][2][8]. The company calls it its first robot foundation model and says one policy runs across tabletop arms, industrial arms and humanoids ^[2]. The claim that sets OM-1 apart from most other generalist robot policies concerns its training data: Reward AI states that OM-1 learns from human demonstrations alone, with no teleoperation and no on-robot experience going into training ^[1].

OM-1 was published as a blog post with embedded videos, not as a paper. At the time of the announcement there were no released weights, no code, no dataset, no API, no success-rate tables and no comparison against public baselines ^[1][8]. Everything described below is what Reward AI reports about its own system.

The Omnibody stack

Reward AI presents OM-1 as the model layer of a larger physical AI stack it calls Omnibody, built around the phrase "One Model, One Data Interface, Any Body" ^[1]. The company's argument is that human-level manipulation will not arrive by scaling data or compute alone, and that capture, learning and control therefore have to be designed together rather than assembled from separately optimized parts. The blog post opens by invoking Philip W. Anderson's 1972 essay "More Is Different" to make that point ^[1].

The hand is treated as the shared interface. Because every demonstration reaches the model in the same form regardless of which robot will eventually execute it, Reward AI argues that data collected now retains its value for hardware that does not exist yet ^[1][4].

Omnibody Hand

Omnibody Hand is a wearable device that a person wears while working, so that the demonstration is recorded without the wearer adapting to any particular robot's kinematics ^[1]. The video labels on the same post also show Omnibody Hands mounted on the robots as end-effectors, so the capture device and the robot hand are the same mechanism, though the blog text does not discuss the robot-side version ^[1]. Reward AI says it builds on DexCap, a portable motion-capture system for dexterous manipulation published by Chen Wang, Haochen Shi, Weizhuo Wang, Ruohan Zhang, Li Fei-Fei and C. Karen Liu at Robotics: Science and Systems in 2024 ^[1][3]. Chen Wang is a co-founder of Reward AI ^[6].

Rather than copying the human hand joint by joint, the design targets three functions the company wanted to preserve: choosing useful contact points, reorienting objects within the hand, and moving between precision and power grasps ^[1]. The result is a seven-degree-of-freedom mechanism. It captures thumb-index pinching along with flexion of the thumb and the index finger, so the thumb can oppose the other fingers as well as the index. For power grasps, the middle, ring and little fingers move together at the metacarpophalangeal joints ^[1].

Reward AI puts unusual weight on fit. Its reasoning is that a device which pinches, slips or mechanically constrains the wearer changes how that person grasps, and a compensated grasp is no longer the behavior the recording was meant to capture. The hand accommodates differences in hand size and finger proportion, and an integrated distal flexion mechanism absorbs differences in finger length, which the company says removes the need for per-user link adjustment ^[1].

One Data Interface

The second layer turns movement into recorded data. Reward AI combines high-frequency tactile feedback, proximity sensing that measures distance before contact is made, and global-shutter in-hand cameras chosen to hold visual context through fast motion ^[1]. The motivating example in the blog post is conveyor-belt sorting, where a person can spot an object, reach, make contact and toss it into a bin inside a fraction of a second, so that losing even a brief window of the interaction makes the recording much less useful. The company says the three channels together cover approach, contact and stable grasp, and that this continuous coverage is what makes sorting at human proficiency learnable at all ^[1].

Hand pose is tracked differently from the usual practice. Visual-inertial tracking, the common default, localizes against a visual update rate that limits how well it follows rapid reversals; the usual workarounds either smooth the trajectory after the fact or ask the demonstrator to slow down, which moves the data away from the quick reaches the system is meant to capture. Reward AI therefore augments visual-inertial tracking with electromagnetic sensing, and says its tracking algorithm compensates for environmental electromagnetic disturbance ^[1].

To quantify the difference the company rigidly mounted both trackers to one structure and moved them between two mechanical stops a known distance apart, so any motion beyond that fixed range counts as overshoot. It reports eight speeds and ten runs per speed ^[1].

The eight speeds span roughly 3 to 67 cm/s, and Reward AI states that its approach reduced mean overshoot error by 60% at high speed, averaged over ten runs per speed ^[1]. Force is recorded along the same trajectory, so a demonstration carries both the path taken and the effort it took to follow it ^[1].

The policy

OM-1 consumes the multimodal streams that Omnibody Hand produces: images, tactile signals, inter-finger proximity and hand-pose trajectories ^[1]. Two design choices are singled out. First, each modality is processed at the native sampling rate of the sensor that produced it rather than being resampled to a common frequency, which the company says preserves high-frequency tactile and motion cues alongside lower-frequency visual context. Second, the policy consumes a temporal history of those streams so it can reason about how contact and task progress evolve ^[1].

The outputs carry motion direction, speed, force and the timing of key events such as grasping and moving. Reward AI describes the inputs and outputs together as a common policy interface for expressing manipulation behavior across robot bodies ^[1]. It also says it developed a new architecture built for efficient inference, so that a rich multimodal history can still be turned into actions fast enough to keep a robot moving at human pace, but the blog post gives no architectural details, parameter count or latency figure ^[1].

Training is described as single-stage. Because every demonstration arrives through One Data Interface in the same form, Reward AI says there is no barrier between pre-training and post-training, nothing has to be re-collected for a new robot, and scaling is a matter of adding human data rather than adding a training stage per robot, task or deployment ^[1][4].

Control layer

The control layer sits beneath the policy and turns its outputs into actuation, covering navigation as well as manipulation on mobile robots. Reward AI trains it with reinforcement learning in simulation to account for velocity- and acceleration-dependent dynamics, external disturbances and system delays, and contrasts it with a classical controller that is pushed off its reference by an unexpected load and never recovers ^[1]. The examples given are pulling open a fully closed refrigerator door without knowing how much it will resist, and lifting delivery boxes of varying weight.

Two scheduling details matter at speed. The control layer runs on its own clock at high frequency, continuing while the policy computes its next actions, so variation in inference latency does not interrupt motion. That independence introduces a seam, because successive action predictions may not join smoothly; the control layer therefore optimizes the transition between successive predictions online, which the company says keeps motion continuous through dynamic behaviors such as tossing and swinging ^[1].

Demonstrated tasks

The announcement video runs 2 minutes 19 seconds and was posted to Reward AI's YouTube channel on 14 September 2026 ^[7]. The blog post embeds clips from it, and the about page states that every clip the company publishes runs at 1x speed and is labeled as such ^[4].

The video description also lists a bartending segment ^[7]. Reward AI's summary claim is that OM-1 "picks up a brand-new task, including challenging dynamics and long horizons, from less than 30 minutes of data", which it attributes to the integration of capture, sensing, learning, inference and control rather than to the policy architecture alone ^[1]. The company's announcement post adds zero-shot generalization to new robots, near-human dexterity and efficiency, and multi-robot collaboration to the list ^[2]. None of these are accompanied by measured success rates.

How it differs from teleoperation-based pipelines

Most generalist manipulation policies are trained on trajectories recorded on robots. The Open X-Embodiment effort assembled a dataset from 22 different robots through a collaboration between 21 institutions, covering 527 skills, and used it to train cross-embodiment RT-X policies ^[10]. Vision-language-action models such as RT-2, Physical Intelligence's pi-zero and pi0.5, and Google DeepMind's Gemini Robotics line follow the same pattern; the pi-zero paper, for instance, describes training on a large and diverse dataset drawn from multiple dexterous robot platforms, including single-arm robots, dual-arm robots and mobile manipulators ^[11]. Binding a dataset to the embodiment that collected it is the cost Reward AI says it avoids by routing everything through a wearable instead ^[1][4].

Learning manipulation from humans rather than robots is not itself new. DexCap and HumanPlus, both from the Stanford groups the founders trained in, learn from human motion rather than robot trajectories, and learning from human demonstrations is an established line in imitation learning ^[3][5][6]. Mobile ALOHA, the other Stanford system on the founders' record, sits on the teleoperation side of that split. Reward AI's claim is stronger: not that human data helps, but that human data suffices, with the robot appearing only at deployment ^[1].

What is not public

The gap between the claims and the public evidence is wide. There is no technical report or peer-reviewed paper, no model weights, no code, no dataset release and no API ^[1][8]. No quantitative task results are given: the only measurements in the post are the tracking-overshoot figures, which characterize the capture hardware rather than the policy. The architecture is described only as "novel" and "built for efficient inference", with no parameter count, context length, action representation or control frequency. The robot platforms used in the videos are not named, and the amount of human data used for training is not stated. Readers who want to weigh the results have the videos, the prose, and the company's word.

The two trade outlets that covered the launch, MarkTechPost on 14 September and Metaverse Post the following day, summarized the blog post and attributed its claims to the company rather than testing them. MarkTechPost noted that no weights, code, dataset or API had been released, so developers could not run the model on their own hardware; Metaverse Post framed the integrated-stack claim conditionally, as something that would matter "if validated" ^[8][9].

Suggested citation

Reward AI asks that the work be cited as: Reward AI Team, "OM-1: Frontier Robot Intelligence, Learned Firsthand from Humans", Reward AI Blog, Sep 2026 ^[1].

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30Reward AI Team, "OM-1: Frontier Robot Intelligence, Learned Firsthand from Humans", Reward AI Blog, September 2026. rewardai.com/...OM-1
2. ^1 ^2 ^3Reward AI (@RewardAI_), announcement post, X, 14 September 2026. x.com/...2099553899804053992
3. ^1 ^2Chen Wang, Haochen Shi, Weizhuo Wang, Ruohan Zhang, Li Fei-Fei, C. Karen Liu, "DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation", arXiv:2403.07788, Robotics: Science and Systems 2024. arxiv.org/...2403.07788
4. ^1 ^2 ^3 ^4"Towards Human-Level Robot Intelligence and Beyond", Reward AI about page. rewardai.com/about
5. ^Zipeng Fu, personal website. zipengfu.github.io
6. ^1 ^2Chen Wang, personal website. chenwangjeremy.net
7. ^1 ^2"OM-1: Frontier Robot Intelligence Learned Firsthand from Humans", Reward AI, YouTube, 14 September 2026. youtube.com/watch
8. ^1 ^2 ^3 ^4Asif Razzaq, "Reward AI Releases OM-1: A Robot Policy Trained on Human Demonstrations Only, With No Teleoperation or On-Robot Data", MarkTechPost, 14 September 2026. marktechpost.com/...teleoperation-or-on-robot-data
9. ^Alisa Davidson, "'One Model, One Data Interface, Any Body': Reward AI's OM-1 Learns Manipulation Straight From Humans", Metaverse Post, 15 September 2026. mpost.io/...arns-manipulation-straight-from-humans
10. ^Open X-Embodiment Collaboration, "Open X-Embodiment: Robotic Learning Datasets and RT-X Models", arXiv:2310.08864, ICRA 2024. arxiv.org/...2310.08864
11. ^Kevin Black et al., "π0: A Vision-Language-Action Flow Model for General Robot Control", arXiv:2410.24164, 2024. arxiv.org/...2410.24164