Skild Brain is a robotics foundation model from Skild AI, founded in 2023 by researchers Deepak Pathak and Abhinav Gupta. Its defining claim is omni-bodied intelligence: rather than being overfitted to one robot, a single model is meant to control radically different hardware — quadrupeds, humanoids, table-top arms and mobile manipulators — even without knowing a robot’s exact body, and to transfer skills learned on one platform to another. Because there is no large-scale ‘internet of robotics’ data, Skild pre-trains the model on alternative sources: large-scale physics simulation (using NVIDIA Isaac Lab and Omniverse) and human videos from the web, treating humans as another robot embodiment; it then post-trains with targeted real-world data for each customer deployment. The model uses a hierarchical architecture that pairs high-level planning with fine low-level control, and is trained end-to-end to apply low forces so it is safer around people. Demonstrations show robots climbing stairs, recovering balance after being pushed, and manipulating cluttered objects. Limitations: Skild Brain is fully proprietary — no weights, code, API or technical paper — so independent verification is limited; published material is company demonstrations and blog posts; and architectural and parameter details are not disclosed. Skild describes its broader ambition as ‘AGI for the real world’.
Skild Brain is Skild AI's 'omni-bodied' robot foundation model -- a single brain designed to control many robot types, from quadrupeds to humanoids to mobile manipulators, covering manipulation, locomotion and navigation.
Skild Brain is a robotics foundation model from Skild AI, founded in 2023 by researchers Deepak Pathak and Abhinav Gupta. Its defining claim is omni-bodied intelligence: rather than being overfitted to one robot, a single model is meant to control radically different hardware — quadrupeds, humanoids, table-top arms and mobile manipulators — even without knowing a robot’s exact body, and to transfer skills learned on one platform to another. Because there is no large-scale ‘internet of robotics’ data, Skild pre-trains the model on alternative sources: large-scale physics simulation (using NVIDIA Isaac Lab and Omniverse) and human videos from the web, treating humans as another robot embodiment; it then post-trains with targeted real-world data for each customer deployment. The model uses a hierarchical architecture that pairs high-level planning with fine low-level control, and is trained end-to-end to apply low forces so it is safer around people. Demonstrations show robots climbing stairs, recovering balance after being pushed, and manipulating cluttered objects. Limitations: Skild Brain is fully proprietary — no weights, code, API or technical paper — so independent verification is limited; published material is company demonstrations and blog posts; and architectural and parameter details are not disclosed. Skild describes its broader ambition as ‘AGI for the real world’.