DK-Port: construction and validation of port autonomous driving simulation environment based on large language models and reinforcement learning

Port environments differ significantly from conventional driving environments, characterized by high vehicle activity, customized roadways, and mixed traffic involving human-operated and autonomous vehicles. To address the lack of port-specific autonomous driving data and improve algorithm generalization while reducing development costs and time, this paper proposes a simulation environment for port autonomous driving called DK-Port. Leveraging embodied intelligence principles, DK-Port provides a realistic and controllable environment for vehicle interactions. This paper incorporates survey-based expert input, employing zero-shot prompting and chain-of-thought reasoning techniques to enable large language models to assist in designing reward functions efficiently. Then, human-machine mixed-traffic scenarios are constructed using complex road data, and adversarial driver models are trained with the PPO reinforcement learning algorithm to identify safety vulnerabilities in autonomous driving systems. Comparative experiments in four typical scenarios, including straight roads and intersections, demonstrate that DK-Port generates diverse adversarial driving behaviors aligned with real port characteristics, such as dangerous overtaking and abrupt lane cutting. Under the premise of ensuring a reasonable distribution of key metrics, the number of lane changes in the straight road scenario is 2.9 times that of the baseline method, and the emergency braking rate in the intersection scenario is increased by 46.7%.