Learning Task Transition from Standing-up to Walking for A Squatted Bipedal Humanoid Robot

Dingsheng Luo

Abstract:

Robots can perform a number of complex tasks,
some of which need to be executed consecutively in an effective
way to form a smooth behavior. In this case, task transition
technique is involved. Due to that there may exist large
differences between two successive tasks, how to switch to the
target task from current robot status smoothly and efficiently
becomes an important problem, especially for those tasks that
are with some strict constraints or necessary requirements, e.g.
stability for a walking task. Unlike previous approaches such
as task weighting and blending, priority handling, kinematic
control, interpolation, etc., in this research, a learning paradigm
based approach is proposed. With respect to the requirements
of efficiency and smoothness, the task transition problem is
formulated as a typical machine learning issue, where those
constraints are taken as factors of an objective function the
learning process based on. Through a specific task transition
problem, which is from standing-up to walking for a squatted
bipedal humanoid robot, our approach is demonstrated and
evaluated. Experimental results on both simulated and hardware
humanoid robot PKU-HR5.1 verify that the proposed
approach is effective.