Dynamic obstacle avoidance strategy and simulation verification of deep reinforcement learning in automatic driving scene

Aiming at the dynamic obstacle avoidance challenge faced by autonomous driving in complex urban environment, this paper proposes a Deep Reinforcement Learning (DRL) framework that integrates multi-modal perception and multi-agent interactive modeling to realize efficient, safe and comfortable end-to-end obstacle avoidance decision. This method models environmental uncertainty using a Partially Observable Markov Decision Process (POMDP), employs the Soft Actor-Critic (SAC) algorithm to train the policy network, and incorporates attention mechanisms and game theory to model other vehicles' behavior, thereby enhancing the multi-agent system's interactive perception capabilities. Through the high-fidelity CARLA simulation platform to build a variety of dynamic scenarios, combined with domain randomization and curriculum learning strategies to improve the generalization ability of strategies. Experimental results show that the proposed method is superior to traditional model predictive control (MPC) and mainstream DRL algorithms (PPO, SAC) in terms of success rate, safety and comfort, and the success rate is still above 85% in unknown scenes, which verifies its robustness and practicability in complex dynamic environment.