Research on Path Planning Algorithm Combining Deep Reinforcement Learning and Theta* Algorithm
Year: 2025
Type: article
Abstract: With the popularization of unmanned systems in practical applications, how to improve their autonomous navigation ability in complex and changeable environments has become a key issue. Based on this, we propose an innovative hybrid path planning method that cleverly combines the Theta* algorithm with deep reinforcement learning technology to achieve more efficient and flexible path planning. The Theta* algorithm is known for its ability to generate smooth and nearly straight global paths in grid environments, making it suitable for tasks that require efficient and direct path planning. However, in dynamic or unknown environments, local path optimization and real - time obstacle avoidance become key challenges. To this end, we introduce deep reinforcement learning technology, which can extract features directly from sensor data through an end - to - end learning mechanism and make optimal decisions based on the current environmental state to achieve effective obstacle avoidance. We first review the Theta* algorithm and its improved versions, such as PS - Theta* and W - Theta*. These algorithms improve the path - finding efficiency while maintaining the path smoothness. Then, for the proximal obstacle avoidance problem, a strategy of integrating deep reinforcement learning into the Theta* framework is proposed. This strategy uses a convolutional neural network (CNN) to process perceptual inputs, such as camera images or LiDAR point cloud data, and adopts DQN and its variants for action selection, thus realizing closed - loop control from perception to action. In addition, this paper also discusses how to balance the relationship among path length, smoothness, and safety through the design of the reward function. Experimental results on CARLA show that the proposed hybrid method not only inherits the advantages of the Theta* algorithm but also exhibits better adaptability and robustness when facing suddenly appearing obstacles. This makes the method particularly suitable for applications in fields such as driverless vehicles and drones, which have high requirements for path smoothness and real - time response. Future work will focus on further optimizing the model training process and exploring the application potential in more types of environments.
Author: Haoyang Qin
Institution: Shandong University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4408870700