Long-horizon motion discovery for humanoid motions from scratch can be challenging without teleoperation or motion retargeting from human demonstrations. MotionDisco is designed to address that by combining a long language model guided “evolutionary search over sequences of interactions with an efficient sequential kinodynamic trajectory optimizer and pruning strategy, enabling the rapid discovery of novel skills. ” With this approach it is possible to discover whole-body trajectories across challenging long-horizon tasks.
Reinforcement learning tracking policies can be trained on these trajectories to transfer motions to a real humanoid robot. The above video shows what this is all about.
[HT]