Efficient-enhanced Reinforcement Learning for Autonomous Driving in Urban Traffic Scenarios

Decision intelligence based on reinforcement learning has gained considerable attention. Compared to the rule-based approach, the reinforcement learning-based approach shows great potential in addressing the challenge of high interaction. However, in urban traffic scenarios, decision-making of autonomous vehicles remains a great challenge, including sample efficiency and stability. In this paper, we develop an efficient reinforcement learning approach with advanced features toward end-to-end navigation in urban traffic scenarios. Firstly, bird's-eye-view (BEV) semantic segmentation is served as the concise representation of traffic scenarios, which can boost the training process. Then the expert demonstration is applied to guide the exploration of the policy at the initial training stage, thereby improving the sample efficiency. At last, we employ the quantile regression to estimate the value distribution in order to improve the stability of the policy. We validate our approach using different simulated traffic scenarios. Experimental results show that our approach has better performance in terms of convergence and stability compared to other baselines.