This AI-Powered Robot Keeps Going Even if You Attack It With a Chainsaw (www.wired.com)

Skild AI unveiled an “omni‑bodied” generalist control algorithm—Skild Brain (and a smaller research variant called LocoFormer)—that’s trained across many different robot morphologies and tasks so a single model can control unfamiliar hardware and rapidly adapt to extreme damage or novel terrain. In demonstrations the system kept robots moving after legs were cut off, had a quadruped stand and walk on its hind legs, balanced a hybrid wheeled‑leg platform after motors were disabled, and generalized from simulated arm data to control real robot manipulators. The team attributes this robustness to large, diverse cross‑robot training rather than single‑platform teleoperation or narrow simulators. The significance for AI/ML is twofold: it showcases a path toward robotics that mirrors the leap in capability seen with large language models—scale and diversity of training yield generalization across form factors—and it introduces an “in‑context” style adaptation in physical control, letting the model rapidly recompose learned skills for new failures. Practically, that could cut deployment time, reduce reliance on per‑robot engineering, and accelerate industrial adoption of adaptive robots (Skild raised $300M in 2024). The work also raises safety and ethical questions as robots become more resilient to physical tampering, putting pressure on standards for control, testing, and governance as generalist robot AI matures.