DOAMC: A Novel Path Planning Algorithm for Dynamic Obstacle Avoidance in Mapless Robot Navigation

Agent navigation without prior maps poses a challenge for traditional algorithms. Although learning-based methods can operate in such mapless scenarios, their performance often degrades significantly when facing the Dynamic Obstacle Avoidance (DOA) problem, manifesting as unreliable avoidance behavior and slow convergence. To address this problem, this paper proposes a novel algorithm (DOAMC) which integrates a Model Predictive Control (MPC) planner into the Soft Actor Critic (SAC) framework. First, the DOAMC algorithm employs Model MPC for local dynamic obstacle avoidance and SAC for higher-level navigation control. Then, to enhance training performance, an improved Prioritized Experience Replay (PER) mechanism is integrated, boosting data utilization efficiency and accelerating convergence. Next, a novel reward function incorporating a potential energy field is carefully shaped to balance safety, efficiency, and path smoothness. Finally, the feasibility of algorithm is validated in a custom Gazebo simulation environment featuring dynamic obstacles. Simulation results demonstrate the capability of robot to successfully complete navigation tasks in complex settings, achieving significantly improved success rates and convergence speeds compared to baseline methods.