Deep Reinforcement Learning Approaches for Motion Planning of Autonomous Buses in Uncertain Environments

Advancements in sensor and communication technologies have accelerated the development of autonomous driving, expanding its application possibilities. The primary challenge is the uncertainty in dynamic urban environments, which affects motion planning algorithms. This paper develops a partially observable markov decision framework tailored for autonomous buses, integrating behavioral decision-making with motion planning. First, we propose an environmental modeling method sensitive to uncertain factors and a specialized hardware framework for autonomous buses. Second, we derived a mathematical model for optimal bus centering on curved roads, informing the reward structure for reinforcement learning. Third, utilizing historical data, we implemented a time-dependent deep reinforcement learning algorithm, recursive deterministic policy gradient (RDPG), to enhance observation accuracy and determine the optimal driving strategy. Our simulations confirm that this algorithm surpasses existing technologies in performance.