Research on Generalization of Autonomous Driving Behavior Policies Based on Policy Constraints

The development of autonomous driving holds significant potential to enhance traffic efficiency and improve societal well-being. However, current autonomous driving systems heavily rely on high-dimensional raw perception data, such as images and LiDAR point clouds, within their decision-making models. This reliance often leads to overfitting to specific training environments, resulting in poor transferability and weak generalization capabilities in new driving scenarios. To address these issues, this paper proposes a framework that incorporates adversarial networks to constrain the encoder's output on top of Proximal Policy Optimization (PPO), combined with sensor fusion(ASF-PPO). This enables the agent to learn key generalized features, enhancing its generalization ability. The proposed method not only enhances the agent's generalization capability in unseen environments but also accelerates the convergence of the training process. We validate the effectiveness of our approach in the CARLA simulation environment. Experimental results show that, compared to Soft Actor-Critic (SAC) and Deep Deterministic Policy Gradient (DDPG), our method achieves significant advantages in both training speed and performance, while also demonstrating excellent generalization on previously unseen road scenarios.