Active Online Learning of the Bipedal Walking

Dingsheng Luo

Abstract:

For legged robot walking pattern learning, the current
mainstream and state-of-the-art researches are most under a socalled
computer simulation based framework, where the walking
pattern is learned via a pre-established simulation platform.
However, when the learned walking pattern is applied to a real
robot, an additional adapting procedure is always required, due
to the big difference between simulation and real walking
circumstances. This turns out to be more critical for a bipedal
walking, because its controlling is more difficult than others, such
as quadruped robot. In this paper, a novel framework for active
online learning bipedal walking directly on a physical robot is
proposed. To let the learning procedure to be of both fast
convergence and high efficiency, a polynomial response surrogate
model, an orthogonal experimental design based active learning
strategy as well as a gradient ascent algorithm are used. The
experimental results on a real humanoid robot PKU-HR3 show
its effectiveness, indicating that the proposed learning framework
is a promising alternative for bipedal walking pattern learning.