Deep Reinforcement Learning-Based Mobile Robot Navigation in Unknown Environments
Benpeng Huang, Jianliang Mao, Chuanlin Zhang · 2025
Efficient and safe autonomous navigation of mobile robots in dynamic and unknown environments remains a central challenge in robotic navigation. Most traditional path-planning methods rely on prior map information and perform poorly in complex, map-less scenarios with multiple moving obstacles and randomized goal positions. Although reinforcement learning can partially address these issues, it still suffers from sparse rewards and low sample efficiency. In this paper, we propose a hybrid navigation framework that integrates velocity obstacles (VO) priors with soft actor-critic (SAC) reinforcement learning to enhance training sample efficiency and convergence speed. At the same time, we leverage the dynamic window approach (DWA) to assist in reward function design, alleviating sparse-reward problems, and employ curriculum learning to guide the agent from simple to complex tasks. We validate our approach in the PyBullet simulation environment, and results show that, compared to both traditional reinforcement learning and classical path-planning algorithms, our method delivers significant improvements in safety, efficiency, and robustness.