Humanoid Gait Optimization Based on Human Data

Abstract

Achieving a stable, human-like gait for humanoid robots is a challenging task. While a variety of techniques exist to generate stable walking patterns, only little attention has been paid to the resemblance to the human gait. Popular gaits, for example, apply the strategy to bend the knees and to swing the torso in the lateral direction in order to ensure stability by shifting the center of mass. As a result, the walking patterns do not look very human-like. However, human resemblance is an important aspect whenever robots are designed to coexist and interact with humans. In this article, we present techniques to optimize a given, stable gait of a humanoid robot with respect to human resemblance. To acquire human data, we use a full-body motion capture system. We propose four different optimization algorithms that work at joint angle basis and use the joint angle difference as measure of similarity. The experiments carried out with a HOAP-2 robot in simulation demonstrate that all techniques generate a gait that is significantly more human-like compared to the robot's initial gate. As the results show, the optimization methods based on hill climbing and policy gradient estimation yield the best performance.