مجتبي نوروزي

Path-following controllers play a crucial role in the performance and safety of autonomous vehicles. One of the most well-known and widely used of these controllers is the PID controller, which can deliver acceptable performance with fixed gains. The appropriate values of the controller's coefficients have been examined using two methods: optimization and Ziegler-Nichols. Reinforcement learning-based controllers, however, have greater potential for delivering adaptive and accurate results thanks to their ability to interact with the environment. The PPO algorithm, which is extensively used in the field of autonomous vehicles, has been employed to implement the controller given the complexity of the problem. There are two major challenges with reinforcement learning-based controllers. The first challenge is the proper tuning of the parameters within the algorithms, and the second challenge is the high fluctuation of the generated control signal. To address these issues, we propose a novel combination of two methods: one involves changes in the actor’s neural network structure and the other incorporates two modifications in the PPO algorithm's loss function, resulting in parameter robustness and reduction in control signal fluctuations. In this study, we have eva‎luated and compared the performance of the PID controller and the reinforcement learning-based controller. Based on the control signal and accuracy of the controllers, initial results showed that the reinforcement learning-based controller delivered poorer results compared to the PID controller with fixed gains and reference speed. On the contrary, the PID controller with fixed gains and speed cannot achieve a highly acceptable path-following accuracy. Moreover, the fixed reference speed may not be suitable for all driving conditions, and therefore, we presented a hybrid approach in which the PID controller is optimized using reinforcement learning for a dynamic and interactive adjustment to its gains and reference speed. In subsequent results, under extremely challenging low-friction road conditions, the fixed-gain PID controller did not perform successfully. However, combining the PID controller with reinforcement learning generated superior results compared to the use of each independently. Not only does this combination show significant improvements in normal road conditions, but also provides suitable performance in challenging road conditions. These results illustrate the potential of hybrid approaches in optimizing control systems for autonomous vehicles.

خودرو خودران , رديابي مسير , يادگيري تقويتي , هينه سازي خط‌مشي نزديك مبدا , كنترل‌كننده تناسبي-مشتق‌گير-انتگرالي

Autonomous vehicle , Path tracking , Reinforcement learning , PPO , PID

mojtaba norouzi

javad poshtan