Research on Autopilot Trajectory Splicing and Reconstruction Based on Hierarchical Reinforcement Machine Learning

Aiming at the multi-objective conflict and spatiotemporal continuity problems in automatic driving trajectory planning under complex dynamic scenes, a trajectory splicing and reconstruction method based on hierarchical reinforcement learning is proposed. By constructing a hierarchical decisionmaking framework with a scene semantic parsing layer, a dynamic behavior prediction layer and a multi-objective optimization layer, we combine spatio-temporal graph convolutional networks to extract the interaction features of traffic participants, and utilize the deep attention mechanism to realize the correlation mapping between local trajectory segments and global paths. A hierarchical Q-network is designed to decouple the long time-series decision-making task, and the feasibility of integrating the augmented Lagrangian method and Lyapunov stability to guarantee the trajectory dynamics is verified. Simulation experiments show that the model has a collision rate of less than <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 1 5}$</tex> collisions/kilometer in high-density traffic scenarios on the CARLA platform, a trajectory splicing success rate of more than 97.6 %, and computational latency controlled within 58 ms, which verifies its comprehensive advantages in safety, real-time and smoothing, and provides an effective solution for automated driving decision-making on complex urban roads.