A body-agnostic robot brain that enables any robot to adapt to damage (www.skild.ai)

Skild AI announced an "omni-bodied" robot brain trained across a simulated multiverse of 100,000 different bodies, producing a single controller that generalizes zero-shot to novel robots it never saw during training. Instead of overfitting to one platform, the model learned strategies that transfer across body morphologies and failure modes, showing rapid online adaptation (seconds) to scenarios like limb loss, locked joints, jammed wheels, added stilts, and heavy payloads. Experiments excluded test robots from the training set; the same model handled all tasks without per-robot fine-tuning. Key technical takeaways: training across extreme body diversity forces the controller to discover robust, generalizable motor primitives rather than memorized trajectories; the system demonstrates in-context learning (using prior trials as prompts) to improve performance within a few attempts. Concrete results include recovery from a 4-DOF limb loss after ~7–8s, walking on three legs after ~2–3s, and switching between rolling and walking gaits when wheels jam. Specialist single-body controllers failed catastrophically in these conditions. The work suggests a practical path toward embodied systems that adapt in real time—an important step for deploying robots reliably in unstructured human environments and a potential building block for broader, physically grounded general intelligence.