Neural network control of a three-link leg

Abstract

Locomotion training of legged machines is a difficult and challenging problem in robotics and artificial intelligence. The work focuses on controlling the running stride of a three link leg. The objective is to control the height, distance and angular momentum of the stride, all of which are determined at takeoff from the ground. We use a CMAC neural network (J.S. Albus, 1975) to control the leg during takeoff. The CMAC network uses local learning and also permits incremental learning. This enables the net to be retrained online to produce the correct signals for locally changed conditions. The CMAC controller trains quickly and generalizes well.