How robots learn to move without watching humans

Getting humanoid robots to perform long, complex tasks like moving multiple objects requires figuring out which contact sequences work—a combinatorially explosive problem. MotionDisco sidesteps human demonstrations or teleoperation by pairing an LLM-guided evolutionary search with a trajectory optimizer to explore interaction sequences efficiently. The method discovered novel whole-body motions across challenging tasks and successfully transferred them to a real humanoid robot, marking the first fully automated discovery and deployment of long-horizon loco-manipulation skills.