Omni-bodied
How it works
The mechanism combines Skild’s descriptions with established academic methods. (1) A massive multi-embodiment training distribution: Skild reports training in a "universe with 100,000 different robots" (procedurally generated morphologies) with aggressive domain randomization, plus internet-video pre-training (including humans treated as a form of robot) and targeted real-world/teleop data. (2) An embodiment-general observation/action representation: instead of a fixed body-specific state/action vector, the policy consumes local per-actuator observations or morphology descriptors, so the same network handles varying degrees of freedom. (3) Morphology conditioning and in-context body inference: the model infers the unknown body from recent sensorimotor history rather than being given its kinematics; LocoFormer uses long-context Transformers whose context spans episode boundaries, enabling test-time adaptation. (4) A hierarchical policy stack (Skild Brain): a low-frequency high-level policy (navigation/manipulation intent) drives a high-frequency low-level policy emitting joint angles/motor torques. (5) In-context adaptation at deploy time: compensating for a lost limb in ~7–8 s, switching to a three-legged gait in 2–3 s, detecting a jammed wheel — on robots excluded from the training set (zero-shot). (6) For manipulation, S1 specifies the task via in-context learning from a single video demonstration instead of language.
Problem solved
Conventional controllers overfit to a single morphology and a single task — every new robot, sensor or even hardware fault requires collecting new data and retraining. Omni-bodied Learning amortizes this cost into one reusable model that exploits positive transfer across embodiments and avoids building a separate policy for each robot.
Components
A large corpus combining procedurally generated simulated morphologies, internet video (including humans) and real-world/teleop data; the basis for cross-body transfer.
Consumes local per-actuator observations or morphology descriptors so a single network handles varying degrees of freedom and sensor types.
Official
A Transformer whose context spans episode boundaries infers the unknown body and compensates for hardware faults without retraining (the LocoFormer mechanism).
Official
A hierarchical stack: a low-frequency high-level policy (intent) drives a high-frequency low-level policy emitting joint angles/motor torques.
How the goal is specified: S1 uses in-context learning from a single video demonstration; other generalist policies use language or a goal image.
Official
Implementation
Transfer across very different bodies (arm vs quadruped) can hurt rather than help; positive transfer is not automatic.
Differing degrees of freedom, sensor types and control rates must be reconciled into a shared representation.
A few data-rich embodiments dominate while rare morphologies underfit.
The 100,000-robot universe is simulated; aggressive domain randomization and real-world post-training are required.
The body must be inferred from sensorimotor history, which demands long context and raises inference cost.
Evolution
Shared Modular Policies ("One Policy to Control Them All", Huang, Mordatch, Pathak, ICML 2020) — a single policy for hoppers, quadrupeds and bipeds.
MetaMorph (A. Gupta, Fan, Ganguli, Fei-Fei, ICLR 2022) — a Transformer tokenizing morphology and generalizing to unseen embodiments.
Open X-Embodiment / RT-X (22 embodiments, 21+ institutions) demonstrates positive cross-robot transfer — the data foundation for generalist policies.
Octo (May 2024), OpenVLA (June 2024) and π0 by Physical Intelligence (Oct 2024) — generalist manipulation policies trained on multi-robot data.
Skild AI (Deepak Pathak, Abhinav Gupta, CMU) announces a $300M Series A ($1.5B valuation) and a "general-purpose brain for robots".
The blog "The case for an omni-bodied robot brain" (24 Sep 2025): training over 100,000 robots and in-context damage adaptation — the brand origin of the term.
LocoFormer (Liu, Pathak, Agarwal) — a generalist locomotion model for unseen legged and wheeled robots with long-context adaptation; the peer-reviewed backing of the Skild brand.
Skild introduces S1, learning tasks from a single in-context video demonstration; a claimed 66% success on unseen tasks vs a 9% baseline (company data).
Hyperparameters (configurable axes)
Long context spanning episode boundaries is the key lever for in-context adaptation in LocoFormer.
Number and diversity of training embodiments; Skild reports ~100,000 distinct robots.
The breadth of physics/sensor randomization needed for sim-to-real transfer to unseen bodies.
A low-frequency high-level policy vs a high-frequency low-level policy in the hierarchical stack.
The share of simulation, internet video and real teleop; Skild quotes ~$1 collection : $3 quality control.
In S1 a single in-context video demonstration replaces — per Skild — roughly 380 post-training examples.
Execution paradigm
A dense network (typically a Transformer); its output is input-dependent through morphology conditioning and in-context adaptation.
No expert routing; behavior is conditioned on the in-context-inferred body rather than on expert gating.
Parallelism
RL/imitation training parallelizes trivially across morphologies and simulation environments on large GPU clusters.
Hardware requirements
Massively parallel simulation RL training and video pre-training require large GPU clusters.
Deployment runs as inference on the robot’s embedded hardware; the concept abstracts over the specific embodiment and Skild exposes skills via an API.