Abstract
Today, an undeniable interest is given to the development of socially intelligentrobotic systems and efficient crowd- and socially-aware navigation strategies. Inthis paper, we introduce a novel crowd-aware navigation algorithm that combinesthe A* path planner with the Double Deep Q-Network (DDQN) method.The algorithm, named Social A*, aims to allow safe navigation for mobile robotsin social dynamics and crowded environments. In order to facilitate future realworldimplementation, a new learning environment compatible with the RobotOperating System (ROS) is developed. This allows expert teleoperation to helptrain the DDQN agent and refine the reward function. We conducted extensivesimulations to compare the performance of Social A* with Socially-attentiveReinforcement Learning (SARL*) and Intention Aware Robot Crowd Navigationwith Attention-Based Interaction Graph (IARL). The obtained simulation resultsdemonstrate that Social A* not only surpasses SARL* and IARL but also showsenhanced performance in handling static obstacles. These results showcase theexcellent crowd-aware navigation performance, the efficiency, and the significantpotential of the algorithm.