Synthesis of Three-Layer Dynamic Binary Neural Networks for Control of Hexapod Walking Robots

Abstract

Three-layer dynamic binary neural networks are characterized by binary connection parameters and signum activation function. Depending on the parameters, the networks can generate various periodic orbits. First, as a basic theory, we give a synthesis method that guarantees storage and global stability of desired periodic orbits. Second, as an engineering application, we synthesize networks that generate periodic orbits corresponding to typical walking patterns of hexapod robots. Presenting an FPGA based hardware, switching of the walking patterns is confirmed experimentally.