A BEHAVIOR-BASED LIBRARY FOR LOCOMOTION CONTROL OF KINEMATICALLY COMPLEX ROBOTS

Controlling the locomotion of kinematically complex robots is a challenging task because different control approaches are needed to operate safely and efficiently in changing environments. This paper presents a graph-based behavior description which allows to dynamically replace behaviors on a robotic system. In the proposed approach, every behavior is represented as a directed graph that can be encoded into a data block which can be saved to or loaded from a behavior library. Since this is not a precompiled module like in other systems, the algorithm and parameters of a behavior can still be adapted online by modifying the data that represents the behavior. Thus, machine learning algorithms can optimize an existing behavior to an unknown situation, e.g., a new environment or a motor failure. With a first implementation, it is shown that the proposed behavior graphs are suited for controlling kinematically complex walking machines.