An Evolvable Hardware Layer for Global and Local Learning of Motor Control in a Hexapod Robot

The use of Genetic Algorithms (GAs) to evolve Continuous Time Recurrent Neural Networks (CTRNNs) for locomotion control of a hexapod robot was demonstrated in previous work. Those controllers were shown to be robust to loss of sensory information and other peripheral variations and their operation was explained through qualitative dynamical systems analysis. Although real robot controllers were fielded, our prior methods contained drawbacks limiting the attractiveness of their implementation as real devices. These included centralization of the evolutionary learning algorithms and the inability of those algorithms to evolve effective controllers without inclusion of a priori architectural knowledge. In this paper, we will introduce a generic neuromorphic Evolvable Hardware (EH) learning architecture and show how it can be used to overcome these previous difficulties. We will also introduce a novel modification to previously reported EH methods that improves search efficacy. We will discuss our previous work, introduce relevant concepts from the emerging field of evolvable hardware as well as our chip, and show how it can be used to deliver effective and practical locomotion control. We will focus especially on the integrated systems abilities to self-heal and reconfigure while in service and will examine centralized, decentralized, and hybrid learning configurations.