Robust Trajectory Planning Method for Complex Traffic Scenarios Based on Adversarial Reinforcement Learning

Existing motion-planning methods for complex traffic suffer from imprecise vehicle-dynamics modeling and weak interaction with surrounding traffic participants, which limits safety and robustness. To address this, an adversary agent with an enlarged acceleration/steering action range is introduced to disturb the protagonist's safe driving, and the two agents are trained against each other within an adversarial reinforcement learning (RARL) framework without modifying the simulation environment; a hierarchical risk map evaluates road and obstacle risks, Gaussian process regression learns the residual uncertainty of the vehicle-dynamics model, and a model predictive controller fuses both to plan interaction-aware trajectories, while a multi-head network with adaptive action noise balances exploration and exploitation to accelerate convergence. Comparative experiments show that the adversarial strategy markedly improves comprehensive driving ability: under changes in speed, vehicle number and driving style, the proposed model achieves better safety, driving stability and efficiency than non-adversarial and conventional baselines.