RL-based Optimisation of Robotic Fish Behaviours

Abstract

The paper presents a reinforcement learning (RL) algorithm for the optimisation of robotic fish behaviours. Six independent parameters are abstracted from the motor controller of a robotic fish and used to parameterize the policy of the reinforcement learning. During the implementation, the sampling results are classified and adaptive evolution steps are adopted. The efficient turning speed of the robotic fish is chosen as the optimal criterion. The simulation results show the good performance of the proposed learning algorithm.