Self-organization of locomotory controllers in robots and animals

摘要

In animals, the Central Pattern Generator (CPG) is essential for achieving coordinated locomotion. In robotics, CPGs have been adapted in various forms for the control of walking machines. In order to implement more complex, adaptive locomotory controllers, we have identified three key issues: (1) adaptation and tuning of the CPG to the robot mechanism, (2) the role of CPGs for trunk flexion in the aquatic and terrestrial gait of vertebrates and its relevance to robotics, and (3) the representation of CPGs for use in robotic work. Two methods of examining adaptation and tuning are presented. The first approach is based on the use of Genetic Algorithms incorporating a novel technique called evolution for tuning a locomotory controller. This approach uses an outside critic to give feedback on the progress of movement. In this case, all experiments are carried out on a real hexapod robot called Rodney. In the second instance, we experiment in simulation with a mechanical model of the lamprey. In this work, we use local, adaptive rules (i.e. self-organization) for the adaptation of the CPG parameters. Our results support the hypothesis that early movement in the embryological stage is instrumental for tuning certain parameters in the CPG. We also examine the importance of trunk flexion in terrestrial locomotion and its relationship to a CPG for trunk flexion in an aquatic environment. We conclude that very little needs to be altered to allow an aquatic CPG to control a terrestrial tetrapod. In this case, experiments were conducted in a tetrapod robot called GEO. Finally, we examined an abstraction of the CPG which we call Adaptive Ring Rules (ARR). These rules act on a phase oscillator and on a motor pattern burst function to create an adaptive and easily understood model of the CPG. ARRs find appropriate use in robotics applications as well as in the modeling of vertebrate locomotion development. The adaptive rules are justified on the bases of an analysis of a reduced Morris-Lecar model of a spiking neuron, and on behavioral observations reported in the literature.