Meta Reinforcement Learning for Autonomous Driving with Rapid Adaptation to Drivers*

The development of autonomous driving is pro-pelling mobility into a new era of innovation. However, existing autonomous driving systems struggle to adapt to various drivers with diverse habits and behaviors, compared with simply catering to a few fixed driving styles. As a result, this paper combines meta-learning and reinforcement learning (RL) to propose an autonomous driving algorithm for personalized decision-making and control, i.e., MetaRL-AD algorithm. MetaRL-AD utilizes the off-policy RL to improve the sample efficiency. In addition, a training cycle sampling method is adopted to improve the stability and generalization, which significantly reduces the time to personalization. It is demonstrated that in an environment with high vehicle density, the method proposed in this paper increases the convergence speed by up to a factor of 8 without any rules or knowledge assistance. The results are available as videos at https://youtu.be/3VVHBz92xnQ.