Frame-By-Frame Motion Retargeting With Self-Collision Avoidance From Diverse Human Demonstrations

Abstract

Human-robot motion retargeting is a complex nonlinear problem, due to heterogeneous kinematic configuration between human and robot. Recent efforts aim to tackle the generalizability of motion retargeting on diverse robots, yet challenges persist in handling unseen human motions with varying scales and the absence of an efficient self-collision avoidance strategy. To address these challenges, we introduce a novel motion retargeting framework capable of generating self-collision-free joint angles from human demonstrations with different scales. The problem of motion retargeting is firstly modeled as a neural latent optimization problem. Then, Skeleton-Normalized Graph Convolutional Network is proposed as the basic component of human motion graph encoder to extract scale-independent deep topological features from human demonstrations as initial latent values. To avoid self-collision, we propose a high-precision self-collision detection network based on graph convolutional network for different robot configurations and structures to penalize self-collision during training of the retargeting network.