J.F. Porter, T. Soule, and E.T. Wolbrecht (USA)
Legged robots, neural networks, central pattern generators, evolutionary algorithm, gait optimization
Presented in this paper is a novel approach to gait generation and optimization in a quadruped robot. The first step of this approach is the creation of a functional, yet sub-optimal manually created walking gait. The gait is then reproduced through error back propagation (EBP) training of a neural network based central pattern generator (CPG) and verified in a dynamic model of a quadruped robot. The CPGs are then seeded into the initial population of an evolutionary algorithm which focuses on further optimization of general locomotion rather than the specific motion of the joints. The results demonstrate the efficacy of the presented approach as an effective and expeditious means of generating fitness function optimal gaits in legged robots.
Important Links:
Go Back
