Skild AI sets out its approach of one "omni-bodied" brain for all robot types

Skild AI/Skild Brain

Skild AI described its approach: one AI brain trained across ~100,000 simulated robot bodies so it must learn general strategies rather than memorise one design. The model controlled robots it had never seen and adapted its gait within seconds if a leg was locked or removed.

• Omni-bodied training on ~100,000 simulated morphologies (quadrupeds, humanoids, arms, mobile bases).
• Prevents overfitting to single body plan; forces learning of generalizable locomotion and manipulation strategies.
• Zero-shot transfer: controls unseen robots; rapid adaptation (seconds) to morphological perturbation (locked leg, removal).
• Thousand-year simulated time per training run enables sample-efficient learning across embodiment space.