GASIL-SAC: A Novel Reinforcement Learning Algorithm for Mapless Navigation of Wheeled Mobile Robots

摘要

Efficient and safe mapless navigation in unknown environments is critical for deploying mobile robots in diverse real-world scenarios. However, existing methods based on deep reinforcement learning (DRL) and imitation learning (IL) often suffer from low sample efficiency and distribution mismatch, respectively. In this paper, we propose GASIL-SAC, a novel reinforcement learning algorithm that combines Generative Adversarial Self-Imitation Learning (GASIL) with Soft Actor-Critic (SAC) for mapless navigation of wheeled mobile robots. To address the issue of overfitting to limited expert data, we introduce a diversity-prioritized positive memory buffer to collect the agent's past good trajectories, ensuring a diverse set of expert demonstrations. We trained a mapless motion planner using the GASIL-SAC algorithm in a 2D simulation environment and conducted comprehensive comparisons with planners based on SAC and Proximal Policy Optimization (PPO) algorithms. The results demonstrate that GASIL-SAC achieves stable performance more rapidly and generalizes better to previously unseen scenarios, highlighting its effectiveness for mapless navigation tasks.