A joint research team from the Technical University of Munich (TUM), Carnegie Mellon University, and New York University has published MotionDisco, a groundbreaking framework that enables humanoid robots to autonomously discover and execute complex, long-horizon loco-manipulation skills without any human demonstrations.
The paper, released on arXiv (2606.06139) on June 7, 2026, addresses one of the most challenging problems in embodied AI: how to generate diverse, contact-rich whole-body motions—such as climbing onto tables while carrying objects, navigating cluttered environments, or picking boxes from under tables—without relying on expensive and time-consuming teleoperation data collection.
How MotionDisco Works
MotionDisco couples two key components: an LLM-guided evolutionary search over sequences of contact interactions, and an efficient sequential kinodynamic trajectory optimizer with a pruning strategy. The LLM proposes mutations to contact plans based on task descriptions and feasibility feedback, while the trajectory optimizer validates whether the proposed motion is physically realizable. This closed-loop process enables rapid discovery of novel, dynamically feasible behaviors across diverse long-horizon tasks.
Real-World Deployment
The researchers went beyond simulation by training reinforcement learning tracking policies on the discovered trajectories and successfully transferring them to a real humanoid robot. This represents the first-ever deployment of long-horizon humanoid loco-manipulation skills discovered entirely through automated evolutionary search—a significant milestone toward robots that can learn physical skills independently.
The framework was developed at TUM's ATARI Lab, led by Prof. Majid Khadiv, with co-supervision from Prof. Cordelia Schmid and collaboration with Prof. Angela Dai (TUM) and Prof. Aaron M. Johnson (CMU).
