Training a Robot to Attend a Person at Specific Locations using Soft Actor-Critic under Simulated Environment

Abstract

We present the application of soft actor-critic (SAC) learning algorithm to train a mobile robot to attend a target person at specific locations inside a Gazebo simulator. Since our previous study confirmed that the appropriate attending position while the target person is standing or walking is at his left or his right side, we design a novel U-shaped reward function behind the target person’s position with respect to the robot’s position. To make the robot can better portray the surroundings, we also propose a novel SAC architecture which employs 1D convolutional neural networks to extract features from laser scans automatically during the training process. Our preliminary experiment result shows that the robot is able to attend the target person at the designed location using our proposed reward function and SAC architecture.