Enhanced Adaptation of Reinforcement Learning Algorithms to Actuator Variability in Autonomous Driving: Error Observation and LSTM Prediction

In the rapidly developing field of autonomous driving, reinforcement learning (RL) is a frequently used approach for developing decision-making control algorithms. However, inevitably, there are differences in the training and testing environments of the models, such as the error characteristics of the actuators, which pose significant challenges to deploying RL algorithms in real-world vehicles. Addressing these challenges is critical to enabling practical applications of RL in autonomous driving systems. This study investigates the performance and transferability of RL algorithms under different actuator conditions. We investigate two methods: RL with online error observation, and RL combined with long short-term memory (LSTM) networks for actuator response prediction. The RL algorithm in the first approach adjusts its policy by observing real-time errors, while the second approach uses LSTM to predict actuator responses, thereby incorporating actuator error characteristics into the training environment during training. Experimental results show that both methods can improve the adaptability of RL models to different actuators, and the RL combined with LSTM method is more effective, especially in the environment of large actuator errors. These findings highlight the importance of accounting for actuator errors during training and suggest that combining deep learning with RL can improve the adaptability and robustness of autonomous driving algorithms. This research lays the foundation for developing reliable and transferable autonomous driving systems capable of navigating diverse actuator environments with enhanced safety and performance.