کنترل شتاب طولی و جانبی خودروی بدون سرنشین با استفاده از یادگیری تقویتی عمیق

Accurate control of longitudinal and lateral acceleration in unmanned vehicles is one of the key challenges in the development of safe and efficient autonomous systems. Given the rapid growth of autonomous technologies and the need to improve the performance of these systems in various road conditions, research in the field of intelligent control of these vehicles has gained special importance. In this paper, an innovative method for controlling the longitudinal and lateral acceleration of unmanned vehicles based on reinforcement learning is presented. The proposed method, by utilizing a disaggregated architecture, reduces the computational complexity and enables the use of more advanced reinforcement learning algorithms. In the initial stage, a single reinforcement learning agent is designed and trained to simultaneously control longitudinal and lateral acceleration. Then, in order to improve efficiency, the reinforcement learning agents are separated into two independent parts for longitudinal and lateral control and trained separately. Simulation results show that this separation not only increases the convergence speed of the training process but also significantly improves the accuracy of the control system's performance. Separating the agents results in a reduction of about 2 hours of deep network training time, a 38.2% reduction in average direction error, and a 10.1% reduction in average distance error. These findings could help develop more advanced and safer control systems for autonomous vehicles and play an important role in improving the safety and efficiency of these technologies.