A segmented motion synthesis method for robotic task-oriented locomotion imitation system

Abstract

Recent research highlights the potential of learning agile robotic locomotion by imitating segmented motion data from humans. However, using single-mode motion data for imitation learning is inefficient for task-specific actions, and motion capture and retargeting processes can be time-consuming. To address these challenges, we propose a motion synthesis framework that combines segmented motions to produce task-specific behaviors characterized by natural movement. Our approach involves three main components: the State Variational Autoencoder (SVAE), the Control Network of Synthesized Motion (SMC-Net), and Critical Joint Constraints (CJC). The SVAE learns motion dynamics from segmented movements and encodes them into a latent space, enabling efficient combination of diverse motions during reinforcement learning. The SMC-Net selects optimal postures from segmented data using Deep Reinforcement Learning (DRL), and its integration with the SVAE’s latent space enhances motion realism. Critical joint constraints are incorporated into the reward to further improve motion quality. Testing on two reach-target-and-reaction tasks with three types of motions demonstrated a 2.6-fold increase in mean rewards and a 1.1-fold reduction in task completion time compared to state-of-the-art baselines using single-mode motions.