Robots Use Natural Language to Self-Adapt to Damage with 81% More Accuracy

Researchers developed a new AI system called ZLIK that allows robots to adapt to structural damage using natural language. By learning from descriptions like "broken wheel" or "cracked chassis," the model can predict the robot's damaged kinodynamics without prior training on that specific failure. In simulations, this zero-shot approach reduced kinodynamic modeling errors by up to 81% and successfully transferred from simulation to real-world, scaled robots.