Application of Soft Actor-Critic Reinforcement Learning to a Search and Rescue Task for Humanoid Robots

This paper proposes a novel maximum entropy based reinforcement learning for dealing with a robotic search and rescue task in a complex enclosed environment. The search and rescue task is described as a Markov Decision Process, under which an auxiliary reward function at multiple stages is designed for the robot and its interaction with the specified environment. A variant of the state-of-art reinforcement learning algorithm, goal-based Soft Actor-Critic (SAC), is developed to train a humanoid robot. Simulation results verify the effectiveness of the proposed goal-based SAC algorithm and its advantages comparing with the prototype of SAC algorithm for the same task.