Path-following Control of Fish-like Robots: A Deep Reinforcement Learning Approach

摘要

In this paper, we propose a deep reinforcement learning (DRL) approach for path-following control of a fish-like robot. The desired path may be a randomly generated Bézier curve. First, to implement the locomotion control of the fish-like robot, we design a modified Central Pattern Generated (CPG) model, using which the fish achieves varied swimming behaviors just by adjusting a single control input. To reduce the reality gap between simulation and the physical system, using the experimental data of the real fish-like robot, we build a surrogate simulation environment, which also well balances the accuracy and the speed of training. Second, for the path-following control, we select the advantage actor-critic (A2C) approach and train the control policy in the surrogate simulation environment with a straight line as the desired path. Then the trained control policy is directly deployed on a physical fish-like robot to follow a randomly generated Bézier curve. The experimental results show that our proposed approach has good practical applicability in view of its efficiency and feasibility in controlling the physical fishlike robot. This work shows a novel and promising way to control biomimetic underwater robots in the real world.