Control Strategy of Continuous Brachiation Including Aerial Phase Based on Reinforcement Learning Including Curiosity

Abstract

This paper describes how to create a controller for continuous brachiation including aerial phase by reinforcement learning. There are no previous research that have achived continuous brachiation including aerial phase. Because of the need for efficient search and robust controller, we used reinforcement learning including curiousity and proposed a branch trimming method using a neural network. We evaluate the performance of learning using the proposed method on a simulator, and then control the actual machine using the learned policy.