Automatic collision avoidance decision making for vessels is a critical challenge in the development of autonomous ships and has become a central point of research in the maritime safety domain. Effective and systematic collision avoidance strategies significantly reduce the risk of vessel collisions, ensuring safe navigation. This study develops a multi-vessel automatic collision avoidance decision-making method based on deep reinforcement learning (DRL) and establishes a vessel behavior decision model. When designing the reward function for continuous action spaces, the criteria of the “Convention on the International Regulations for Preventing Collisions at Sea” (COLREGs) were adhered to, taking into account the vessel’s collision risk under various encounter situations, real-world navigation practices, and navigational complexities. Furthermore, to enable the algorithm to precisely differentiate between collision avoidance and the navigation resumption phase in varied vessel encounter situations, this paper incorporated “collision avoidance decision making” and “course recovery decision making” as state parameters in the state set design, from which the respective objective functions were defined. To further enhance the algorithm’s performance, techniques such as behavior cloning, residual networks, and CPU-GPU dual-core parallel processing modules were integrated. Through simulation experiments in the enhanced Imazu training environment, the practicality of the method, taking into account the effects of wind and ocean currents, was corroborated. The results demonstrate that the proposed algorithm can perform effective collision avoidance decision making in a range of vessel encounter situations, indicating its efficiency and robust generalization capabilities.

With the development of autonomous driving technology, the application scenarios for mobile robots and unmanned vehicles are gradually expanding from simple structured environments to complex unstructured scenes. In unstructured three-dimensional spaces such as urban environments, traditional two-dimensional map construction and path planning techniques struggle to effectively plan accurate paths. To address this, this paper proposes a method of constructing a map and planning a route based on three-dimensional spatial representation. This method utilizes point cloud semantic segmentation to extract navigable space information from environmental point cloud data and employs voxelization techniques to generate a voxel map. Building on this, the paper combines a variable search neighborhood A* algorithm with a road-edge-detection-based path adjustment to generate optimal paths between two points on the map, ensuring that the paths are both short and capable of effectively avoiding obstacles. Our experimental results in multi-story parking garages show that the proposed method effectively avoids narrow areas that are difficult for vehicles to pass through, increasing the average edge distance of the path by 83.3% and significantly enhancing path safety and feasibility.