Hierarchical Reinforcement Learning-Based Policy Switching Towards Multi-Scenarios Autonomous Driving

Reinforcement learning has gained a lot of attention and applications in the field of autonomous driving in recent years. However, the actual scenarios of automatic driving applications are often diverse, so the reinforcement learning algorithm using only a single driving strategy is difficult to meet the multiple requirements of efficiency and safety in the multi-scenarios autonomous driving task. To solve this challenge, we propose a hierarchical reinforcement learning structure to learn a unified top-level switching master policy between different driving styles policies. The whole framework uses a bottom-up training manner with diverse reward function designing. Through experimental comparison, our method exceeds the performance of single policy and rule-based switching strategy. Based on this framework, we won the first place in the DAI 2020 Autonomous Driving Workshop single-agent track competition.