Neural control of locomotion in a quadrupedal robot

The authors present results of a first study demonstrating that the apparently complex task of controlling walking in a real quadrupedal robot with highly nonlinear interactions between the control elements can be learned quickly by a crude and simple reinforcement learning algorithm. They can as yet say little that is useful about the contribution of reflexes to learned walking, and nothing about the quality of evolved solutions other than that their discovery by applying genetic algorithms to real robots is likely to take a prohibitively long time. However, they hope that their experiences will point the way to more controlled studies of the applications of reinforcement learning to real-world problems, especially to control problems associated with autonomous mobile robots