DIH2H: Decomposed Imitation for Humanoid Locomotion From Human to Humanoid

In recent years, humanoid robots with flexible locomotion capabilities have attracted significant attention. Imitation from human motion has proven to be an effective approach for them to learn. This article presents a novel training framework called Decomposed Imitation from Human to Humanoid (DIH2H). We define a “key skill set” to represent essential human motions. By introducing an action redirection mechanism, an agent learns joint motion patterns rather than directly copying human motions. This focused approach allows a humanoid robot to effectively handle the differences between its own locomotion and human gaits, thereby preventing unnatural motion. The control policy is trained separately for the robot’s upper and lower body. The lower body employs imitation learning (IL), while the upper one leverages a combination of IL and curriculum learning. By gradually introducing increasingly challenging tasks and diverse environments, the upper body can robustly adapt to dynamic changes across various terrains and motions. Simulation results demonstrate that DIH2H improves policy learning efficiency and generalization across diverse tasks and environments, achieving a 10% increase in imitation accuracy and 25% higher data efficiency over its existing peers.