Heuristic reward-based Soft Actor-Critic method for autonomous mobile robot navigation

This paper proposes a study on heuristic-based DRL techniques for autonomous mobile robot navigation to address this performance degradation. The Soft Actor-Critic (SAC) method is considered for implementation due to its ability to balance exploration and exploitation, efficiently reusing past experiences to accelerate learning and reduce sampling complexity. The proposed heuristic SAC (h-SAC) system utilizes inputs from the neural network, including previous linear and angular velocities, ten 2D LiDAR data points, and the angle and relative position of the mobile robot with respect to the goal. A goal-based heuristic reward function is proposed to accelerate exploration by providing a positive
reward when the goal is reached and a negative reward when obstacles are encountered. The proposed h-SAC method is implemented in the Robot Operating System (ROS)