Latest / Embodied AI 101 / OM-1: Human Skills, Many Bodies

Transcript

OM-1: Human Skills, Many Bodies · Embodied AI 101

• OM-1: Human Skills, Many Bodies. OM-1: Frontier Robot Intelligence, Learned Firsthand from Humans • Reward AI Team • Reward AI • Company technical blog • 2026. The important claim is about where robot learning begins. Imagine collecting a manipulation demonstration before choosing the robot that will execute it. Someone performs a task, the recording enters a shared dataset, and a future machine learns from that example without requiring the person to demonstrate again through its particular joints, gripper, or teleoperation interface. That would change more than the cost of collecting demonstrations. It would change what a demonstration represents. Instead of recording how one machine accomplished a task, we would record something closer to the underlying interaction: where contact happened, how an object moved, and what the demonstrator did when the world resisted. This is the useful lens for understanding Reward AI’s Omnibody effort. The company describes a common data interface and policy spanning robot arms, legged humanoids, and wheeled mobile manipulators, rather than a separate demonstration-collection process for each deployment. One correction to the headline is essential: OM-1 is described as learning from instrumented human manipulation without teleoperation or on-robot experience—not from ordinary human video alone. Those are substantially different propositions. The interesting question is not simply whether a sufficiently large network can watch people and discover robot control. It is whether we can design a measurement and execution interface that makes human demonstrations usable across machines. As of September 15, 2026, the material I could verify consists of a company technical announcement and demonstrations, rather than a public OM-1 research paper or downloadable implementation. This episode therefore analyzes a disclosed system direction, not a reproducible benchmark submission. My reading is that the strongest contribution to investigate is the integration boundary. What should the human demonstrate? What should the policy predict? What must the robot controller guarantee? And which aspects of embodiment can remain below that boundary?. We will separate the company’s disclosures from our own interpretation, connect the approach to closely related research, and finish with the experiments that would make its generalization claims much more informative. A demonstration already contains an embodiment. Consider two hypothetical recordings of the same task: putting a lid on a container. In the first, a person teleoperates a robot. The trajectory includes pauses caused by the interface, corrections to compensate for camera placement, and perhaps a different grasp because the robot cannot approach from the person’s preferred direction. In the second, the person handles the container directly. Now the demonstration contains more natural coordination—but its motions may depend on capabilities the target robot does not possess. Neither recording is automatically the right training example. The first makes execution compatibility easier to establish. The second potentially captures a better strategy, while creating a harder correspondence problem. Universal Manipulation Interface, or UMI, provides an important precedent. Cheng Chi and colleagues used handheld grippers with wrist-mounted cameras to collect demonstrations without operating a robot. Their design combined comparable observations during teaching and deployment with relative action trajectories and latency matching. The interface was engineered to make the resulting behavior transferable, rather than expecting a policy to infer every correspondence from unrelated human and robot observations. That suggests a useful distinction: collecting data without a robot does not require collecting data without constraints. A carefully selected constraint can make the data substantially more useful. For example, imagine standardizing the location of the observation camera relative to the contact surface. That sacrifices some freedom in the capture device, but could remove a recurring visual mismatch from every training example. Whether that trade is worthwhile depends on how much behavior the device prevents and how much ambiguity it eliminates. A different strategy appears in MimicPlay, from Chen Wang and colleagues. There, human play data supplies high-level latent plans, while a smaller set of teleoperated demonstrations trains the robot’s low-level visuomotor behavior. Human and robot data play complementary roles rather than sharing a fully interchangeable action interface. These approaches expose the real design space. We can ask human data to teach task structure, teach executable movements, or supply both. Each choice creates different requirements for sensing, correspondence, and control. For OM-1, I would therefore evaluate the data interface as seriously as the neural network. A portable policy is only as portable as the meaning of its observations and actions. Standardizing a file format is easy compared with standardizing what a recorded movement means physically. The wearable is part of the learning algorithm. Omnibody Hand has seven degrees of freedom, emphasizing thumb–index manipulation and coupled middle, ring, and little-finger grasps. The capture system records cameras, touch, proximity, force, and hand motion, with electromagnetic sensing augmenting visual-inertial tracking. This belongs to a recognizable research lineage. DexCap, introduced by Chen Wang and colleagues at RSS 2024, combined portable hand-motion capture with environmental observations. Its DexIL pipeline retargeted motions through fingertip inverse kinematics and trained a point-cloud-based diffusion policy. DexCap also included an optional human-in-the-loop correction mechanism; it should not be retroactively described as having exactly OM-1’s claimed training restrictions. The design question I find most interesting is how much hand complexity an interface actually needs. Suppose we are building a collection device for packaging tasks. We might care deeply about opposing a thumb against another finger, stabilizing a box, and changing between a pinch and a broader grasp. We might care less about reproducing every independently articulated motion of a biological hand. That would be a task-distribution hypothesis, not an anatomical claim. The device would be betting that a smaller set of controllable contact configurations preserves most of the useful behavior in its intended domain. Classical grasp analysis gives a reason to think in these terms: the ability to resist object disturbances depends on contact locations, contact normals, friction, and available forces—not merely on how many joints a hand contains. Even a geometrically suitable grasp may fail if its actuators cannot generate the required forces. But reducing dimensionality can also hide exclusions. A compact interface might represent a broad family of useful grasps while poorly capturing independent finger repositioning or particular tool interactions. I would want its evaluation organized around these functional boundaries, rather than treating the degree-of-freedom count as a capability score. I would also test the demonstrator, not just the tracking system. Does wearing the device change the chosen grasp? Do people avoid narrow spaces? Do they replace an in-hand adjustment with a table-assisted maneuver? Those changes might be perfectly acceptable, but they should be measured. This leads to a practical data-quality metric I would propose: the fraction of natural task strategies preserved by the capture interface. Tracking accuracy measures how faithfully we recorded the performed behavior. It does not tell us whether the apparatus caused the person to perform a different behavior in the first place. The wearable, in that sense, is not just an input peripheral. Its mechanical choices shape the learning problem. What a common policy interface must preserve. Reward AI describes single-stage training, native-rate multimodal histories, and actions specifying motion, speed, force, and event timing. Architecture, parameter count, dataset scale, and task-success tables remain undisclosed. That is enough to discuss the interface, but not enough to reconstruct the model. We should not silently fill the gap with our favorite diffusion transformer, flow-matching architecture, or vision-language backbone. Consider a hypothetical grasp during rapid sorting. Before contact, a finger is approaching the object. Shortly afterward, it is touching the object but has not secured it. Later, the object is supported well enough to accelerate away. Those moments might look similar in an image while requiring different actions. A useful representation should distinguish them without forcing the policy to reconstruct every detail from appearance alone. Contact-mode analysis makes the underlying distinction explicit: rolling, sliding, and breaking contact impose different constraints on feasible motion. This is why “the fingers are near the object” is not a complete description of manipulation state. For an implementation inspired by OM-1, I would preserve accurate timestamps before deciding how to fuse sensor streams. Keeping a high-rate signal is valuable only if its timing relative to observations and commanded actions remains meaningful. For instance, I would test two failure cases separately. In one, a tactile event is absent. In the other, the event is present but shifted in time. These could produce different learned mistakes, even if the tensors have identical shapes and approximately similar statistics. The deployment contract deserves equally careful treatment. What supplies the corresponding sensory channels on the robot? Are forces expressed in physical units or normalized coordinates? What happens when a channel is unavailable? Does the policy know the achievable range of the receiving machine?. These are questions for a future specification, not assumptions we can infer from a block diagram. The training description also leaves a provenance question. A single task-policy training stage does not, by itself, establish how every encoder was initialized. Nor does it tell us whether adding a new skill means retraining on a pooled dataset, continuing optimization, or some other procedure. For researchers, the essential deliverable would be an explicit observation-and-action contract, including timing, coordinate frames, calibration requirements, and missing-data behavior. Without that, “one interface” remains a systems aspiration rather than something another laboratory can implement faithfully. The controller explains how transfer could work. The execution layer uses simulation-trained reinforcement learning, runs independently of policy inference, and optimizes transitions between successive predictions. This is the crucial qualification to the human-only story. The manipulation demonstrations may be human-only while the execution machinery still learns from simulated robots. There is no contradiction. It is a division of responsibility. UMI on Legs, by Huy Ha and colleagues, demonstrated a closely related separation. A manipulation policy predicted end-effector trajectories, while a simulation-trained whole-body controller executed them on a quadruped with an arm. The work also transferred an existing manipulation-policy checkpoint originally intended for a fixed-base arm. Its interface used task-frame trajectories, allowing the controller to compensate for body motion while preserving the desired manipulation behavior. The important insight is that a task policy need not directly decide every leg or arm joint command. A lower-level system can take responsibility for realizing the requested movement on a particular body. HumanPlus provides another useful comparison. Zipeng Fu and colleagues trained a low-level humanoid controller in simulation using human motion data. Human shadowing then enabled real-world teleoperation, and the resulting robot demonstrations trained autonomous skill policies. That differs from dispensing with on-robot task demonstrations, but illustrates why controller learning and skill learning need separate descriptions. For OM-1, I would not assume that a shared task-policy checkpoint implies identical controller weights everywhere. A common policy could coexist with different controllers, calibration procedures, or robot descriptions. Establishing the boundary would make the claim clearer, not less interesting. Now imagine moving the same hand trajectory between a fixed arm and a mobile robot. The fixed arm may execute it without changing its support configuration. The mobile system might need to reposition its base or brace before making contact. Our proposed abstraction succeeds if both systems can satisfy the same task-level request. It fails if the request omits information needed to decide how to execute it. Force is particularly important here. Hybrid motion-force control distinguishes motion along unconstrained directions from force along constrained directions. A request to interact with a surface is not generally interchangeable with a request to occupy a Cartesian position. I would consequently resist interpreting any favorable learned-controller demonstration as evidence that classical control cannot handle contact or disturbances. The meaningful comparison is between specified implementations under matched conditions. Finally, no interface removes physical feasibility. A receiving robot still needs adequate reach, contact geometry, force capacity, and support. The useful ambition is reusable skill knowledge within an explicit execution envelope—not a policy that makes every conceivable body capable of every demonstrated action. Speed is an end-to-end property. A common mistake in interpreting fast robot behavior is to look for a single inference-frequency number. For a system like this, I would instead ask when the observation was captured, when the prediction became available, which part of the prediction was executed, and how the controller responded between predictions. Those timestamps describe the actual feedback loop. Real-Time Execution of Action Chunking Flow Policies, by Kevin Black, Manuel Galliker, and Sergey Levine, addresses this problem directly. Their real-time chunking method generates a new action sequence while the current one executes, preserving an execution prefix and completing the remaining trajectory through inpainting. The work shows why simply predicting longer action chunks does not automatically solve latency and discontinuity problems. This is a relevant comparison, not evidence that OM-1 uses the same algorithm. For our own implementation, I would separate continuity from responsiveness. Continuity asks whether commands join smoothly. Responsiveness asks whether the robot can change its behavior when a new observation invalidates the plan. A system could score well on one and poorly on the other. Imagine a smooth reaching motion toward an object that someone removes. Blending successive predictions beautifully is not sufficient if the execution system remains committed to a stale contact event. I would therefore test disturbances at specific phases: during approach, immediately before contact, during grasp stabilization, and after lifting. The policy might require different interruption behavior in each case. There is also a reason to avoid treating slow playback as an adequate proxy for fast execution. Dynamic manipulation involves inertial effects that quasistatic reasoning intentionally neglects. Changing the timing can change the physical problem, rather than merely changing how quickly the same problem is solved. The engineering lesson I take from this family of work is to optimize an observation-to-action system, not just a network forward pass. For high-speed evaluation, I would report latency distributions and action age alongside inference throughput—and inspect what happens during the worst delays, not only the median. “Zero-shot” needs a held-out axis. The phrase zero-shot is useful only when we say what was absent before evaluation. For this episode, I would distinguish four meanings. First is zero target-robot task demonstrations. The high-level policy receives no manipulation examples collected through the target machine. This can still be compatible with a calibrated robot, a known kinematic model, and a trained low-level controller. Second is zero-shot embodiment transfer. A frozen task policy executes on a different body. To make this informative, the experiment should state what changed: arm geometry, mobility, hand mechanism, sensing, controller, or some combination. Third is zero-shot task generalization. The system performs a task that was not represented in its task-training data. That is a different claim from transferring a familiar skill to new hardware. Fourth is zero-shot environment or object generalization. The robot encounters unfamiliar layouts or objects without adaptation. Again, that says nothing by itself about whether the skill or body is new. Reward AI reports learning a new task from less than thirty minutes of human data. That could be an important sample-efficiency result. It is not, however, zero examples of that task. Nor does thirty minutes of collected data tell us the training duration, the total historical training budget, or the amount of prior overlap with the new skill. These distinctions should not be treated as semantic objections. They identify which part of the scalability problem has improved. Suppose a new skill requires a small human dataset, after which a frozen policy deploys across several compatible machines. That would combine few-shot task acquisition with zero-shot embodiment transfer. It could be extremely valuable even though it is not zero-shot along every axis. My preferred evaluation would freeze the policy first, document the permitted integration work, and then introduce the held-out condition. Otherwise, adjustments made while watching test rollouts can quietly become part of the learning procedure. The most useful question is therefore not “Is OM-1 zero-shot?” It is: What is held out, what remains frozen, and what work is permitted before the first scored trial?. What the evidence establishes—and what it does not. In a mechanical-stop tracking test—eight speeds, ten runs each—the company reports mean overshoot falling from 24.9 to 9.5 millimeters at the highest speed, roughly sixty percent, versus visual-inertial tracking. That is a component measurement with a defined protocol. It is also easy to overinterpret. The experiment measures overshoot against a known travel boundary. It does not directly measure robot fingertip accuracy, policy success, orientation error, or force regulation during manipulation. I would read it as evidence that the capture subsystem deserves attention, then ask for complementary tests: full-trajectory ground truth, rotational accuracy, latency, calibration stability, and sensitivity to the environments where demonstrations are collected. The website labels its showcase footage autonomous and at normal playback speed. That is useful presentation information, but not a task-reliability estimate. For the near-human-performance claim, I would want a defined human comparison. Which people? Using which tools? With what task setup and completion criterion? Are failed attempts and recovery time included?. A robot that matches a person’s movement speed on successful trials might still have very different effective throughput. Conversely, a slower robot could be useful if it requires little supervision and operates consistently. The relevant measure should match the intended application. A demonstration can establish that a behavior is achievable under the shown conditions. It cannot, without an evaluation protocol, tell us how often that behavior occurs or how broadly the conditions can vary. Open X-Embodiment offers a useful example of a different evidentiary structure. The original work assembled data spanning twenty-two robot embodiments and evaluated its model questions through 3,600 trials across six robots. Those numbers do not make it a directly comparable benchmark for OM-1; they illustrate the distinction between showing breadth and measuring transfer. For a foundation-model claim, I would additionally ask whether more diverse training data improves held-out performance, whether one shared policy competes with task specialists, and how much new-task data is saved by prior training. Those are the experiments that would distinguish a broad reusable model from an impressive collection of behaviors supported by a common system. The present evidence warrants technical interest. It does not yet justify assigning a general success rate, a human-equivalence score, or a ranking against other foundation policies. One checkpoint is not automatically a shared mind. Reward AI also describes work beginning with quadmanual coordination involving four robots, with a broader ambition for mixed teams that exchange objects, divide roles, and recover together. This opens a separate research question. Running the same weights on several robots does not, by itself, give them shared observations, shared memory, or a common estimate of task progress. We should distinguish a centralized policy producing joint actions from decentralized instances that happen to use the same parameters. Multi-agent learning has long made related distinctions. In Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments, Ryan Lowe and colleagues explicitly incorporate other agents’ information during training to support coordinated behavior. That is not a proposed description of OM-1; it is a reminder that coordination needs an information structure. Consider a hypothetical handoff. One robot must decide when the recipient has established sufficient support to justify release. An identical policy checkpoint does not tell the giver what the recipient currently senses. A centralized observation could resolve that uncertainty. So could an explicit communication signal or a sufficiently informative shared view. Which solution is used matters for latency, scaling, and failure handling. I would evaluate collaboration with interventions designed to break a fixed routine: delay one participant, alter the handoff location, change the payload, or remove a robot’s expected contribution. The decisive question would be whether the team changes its behavior coherently, rather than whether several robots can move in a coordinated-looking sequence. That would also clarify whether collaboration emerges from reusable manipulation knowledge, requires separately collected joint demonstrations, or relies on another supervisory component. All three are legitimate engineering choices, but they imply different scaling stories. How I would investigate the idea in a laboratory. The public research surrounding OM-1 already offers useful starting points without requiring us to pretend we can reproduce its undisclosed internals. DexCap’s repository includes collection, processing, dataset construction, and policy-training code. UMI provides a capture-to-training-to-deployment workflow. UMI on Legs releases simulation-trained whole-body-control and deployment components. These are independently inspectable resources, not substitutes for OM-1 weights. I would begin with a deliberately narrow experiment: one manipulation skill, two robot bodies, and one fixed observation-and-action contract. The first task would be to define the contract in enough detail that two engineers could implement it independently. I would include coordinate frames, units, timestamp semantics, calibration steps, achievable command ranges, and behavior when observations arrive late. Then I would establish an execution baseline before training a task policy. Feed both robots the same family of reference motions and measure their ability to execute them under varying loads and disturbances. If the controllers cannot satisfy the common interface, policy transfer will be difficult to interpret. Next, I would collect human demonstrations with all available signals, while designing the recording process so that modalities can later be removed independently. That permits comparisons between vision-only, vision-plus-motion, and contact-rich observations without changing the demonstrators or task distribution. I would pay particular attention to timing ablations. Remove a signal, delay it, or compress its temporal resolution. These tests answer different questions: whether the information matters, whether its alignment matters, and whether its bandwidth matters. For the learning comparison, I would hold the policy architecture and optimization budget fixed initially. Otherwise, better performance from the instrumented dataset might be confounded with a different model or a more favorable training procedure. For embodiment transfer, I would keep a strict integration ledger. Record which components were frozen, which controller was used, which calibrations changed, and whether anyone adjusted the system after observing task failures. Integration is not disqualifying; hidden integration makes the result ambiguous. I would also split data by collection session, person, object instance, and environment—not merely by adjacent clips from the same recording. The objective would be to construct tests that actually challenge the intended transfer mechanism. Only after establishing this small experiment would I expand the task set. I would choose tasks that stress different contact requirements rather than accumulating visually different pick-and-place examples. Finally, I would measure the total effort required to add a robot and a skill. Human demonstration time is one useful number. I would also record usable-data yield, calibration time, controller-development effort, intervention frequency, and time spent diagnosing failed transfers. This is the practical meaning of scalability I would want to test: does the common interface reduce the marginal work required for the next deployment?. A successful small study could answer that question more convincingly than a much larger demonstration collection with poorly documented boundaries. The takeaway: remove robot demonstrations, not robotics. OM-1 is interesting because it invites us to reconsider the unit of transferable robot knowledge. The strongest version of the idea is not that embodiment ceases to matter. It is that we can choose a boundary where useful interaction knowledge becomes reusable, while sensing and control make that boundary physically meaningful. The surrounding literature already supplies important pieces: robot-free demonstration capture, human-motion retargeting, hierarchical use of human data, and simulation-trained execution layers. My interpretation is that the research opportunity lies in making those pieces work together across a substantially broader distribution. What would most strengthen the OM-1 case is a held-out-robot evaluation with a frozen policy, an explicit deployment contract, repeatable task results, and an accounting of the work required to integrate each machine. Until then, I would treat it as a compelling systems direction rather than a settled demonstration of universal or human-level robot intelligence. For roboticists, the immediate lesson is concrete: before asking a model to overcome an embodiment gap, ask which parts of that gap can be removed—or at least made explicit—through interface design. Robot-free demonstration collection could be transformative. It does not make the physical details disappear. It makes choosing the right physical details even more important.