A Learning Strategy for the Control of a One-Legged Hopping Robot

摘要

We study neural network strategies for the control of a dynamic, locomotive system using as a model of a one-legged hopping robot. The control task is to make corrections to the motion of the robot that serve to maintain a fixed level of energy (and minimize energy losses), which yields a stable periodic limit cycle in the system's state space corresponding to periodic hopping to a prespecified height. The studied models are Michie and Chambers' BOXES system (1962), the ASE/ACE configuration of Barto and his coworkers (1983), and Anderson/Sutton's two-layered Connectionist model (1986.) Results are demonstrated through numerical simulations, and quantitatively compared to performance obtained by Raibert (1984) for the robotic leg, using full-state feefback. The main difference between Raibert's solution and the `neural' strategies presented here is that our system is not aware of the dynamical model of the plant which it is to control. It has to discover how to control the plant through a long sequence of trial and error experiments.