The decision-making performance of autonomous vehicles tends to be unstable at unsignalized intersections, making it difficult for them to make optimal decisions. We propose a decision-making model based on the Variable-Noise Multi-Agent Deep Deterministic Policy Gradient (VN-MADDPG) algorithm to address these issues. The variable-noise mechanism reduces noise dynamically, enabling the agent to utilize the learned policy more effectively to complete tasks. This significantly improves the stability of the decision-making model in making optimal decisions. The importance sampling module addresses the inconsistency between outdated experience in the replay buffer and current environmental features. This enhances the model’s learning efficiency and improves the robustness of the decision-making model. Experimental results on the CARLA simulation platform show that the success rate of decision making at unsignalized intersections by autonomous vehicles has significantly increased, and the pass time has been reduced. The decision-making model based on the VN-MADDPG algorithm demonstrates stable and excellent decision-making performance.

Highway ramp merging requires autonomous vehicles to make safe and efficient decisions in dense mixed traffic, where strong vehicle interactions and rapidly changing acceptable gaps make the task particularly challenging. Existing reinforcement learning methods are often unimodal and overly conservative, while diffusion-based policies, despite their ability to generate multimodal actions, usually suffer from high inference latency and safety risks caused by unconstrained sampling. To address these issues, this paper proposes AM-DIMPO, an action-mask-guided safe diffusion-implicit policy optimization framework for ramp-merging tasks. The proposed method combines DDIM-based implicit sampling with a state-dependent continuous action mask to improve multimodal action generation efficiency while enhancing action feasibility. In addition, the mask correction signal is incorporated into policy learning to encourage the policy to generate actions closer to the safe feasible region. Experiments are conducted in a Gym-based ramp-merging simulator under both light-traffic and dense-traffic scenarios, where the proposed method is compared with classical reinforcement learning baselines, diffusion reinforcement learning baselines, and a safety-aware PPO baseline. The results show that, in dense traffic, AM-DIMPO achieves a merging success rate of 97.3%, an average speed of 16.27 m/s, and an inference latency of 68 ms; in light traffic, the success rate reaches 98.1%. Moreover, the proposed method maintains robust performance under the tested noisy-observation and reduced-visibility settings. Overall, AM-DIMPO achieves a favorable balance among empirical safety, traffic efficiency, robustness, and real-time inference performance in dense highway ramp-merging tasks.