Goal-guided greedy experience replay-enhanced reinforcement learning for efficient autonomous navigation

Despite some success in mapless goal-driven navigation using deep reinforcement learning, there is an issue of insufficient experience utilization in deep reinforcement learning-based mapless goal-driven navigation. The reason is that during experience sampling, the differences between experiences are not fully considered. Uniform sampling leads to the underutilization of experiences that are more beneficial for agent learning. Current deep reinforcement learning-based mapless goal-driven navigation approaches fail to adequately account for this, resulting in low experience utilization efficiency during the agent's learning process. To address this issue, we propose a Goal-guided Greedy Experience Replay Enhanced Reinforcement Learning (GER-RL) method for efficient autonomous navigation. More specifically, we prioritize experiences by their importance, employing non-uniform sampling, and incorporate this experience sampling approach into the reinforcement learning-based navigation model to improve data utilization efficiency. Experiments conducted in a simulation environment show that our method prioritizes experiences that are more useful to the agent, improving the data efficiency of the DRL learning process and significantly enhancing navigation performance. Compared to existing deep reinforcement learning-based methods for mapless goal-driven navigation, our approach demonstrates significant improvements across key performance metrics, including average reward, average length, average time, success rate and collision rate.