Actor-critic design using echo state networks in a simulated quadruped robot

Abstract

In recent years, several studies proposed the application of echo state networks (ESN) to adaptive reinforcement learning schemes for the control of artificial autonomous agents. Especially the actor-critic design (ACD) is a promising candidate for robotic systems with continuous state and action spaces, as was demonstrated in several studies using simple wheeled robots. In the present work, we investigate applicability of this learning framework to more complex robotic systems, namely a quadruped running robot with rich dynamics. New challenges and questions arise, such as the nontrivial mapping of actions to the resulting behavior.