Reinforcement Learning-Based Adaptive Safety Tracking Controller for Autonomous Mining Trucks

Efficient path tracking for autonomous mining trucks in open-pit mines is challenging due to high curvature paths, rapidly changing curvature rates, long control delays, and low control precision. To tackle it, reinforcement learning (RL) was introduced for path tracking control of autonomous mining trucks, through designing separate RL control networks and adjustment networks to adapt to mining paths and the truck's control response. To address the computational time and black-box issues associated with RL, an adaptive preview controller was used to simplify the state inputs and improve computational efficiency, and a safety control strategy was implemented to mitigate the safety risks posed by the black-box nature of the model. Simulation tests on actual mining paths show the RL controller achieves an average lateral error under 0.05 m with an average execution time under 10 ms, and the safety layer effectively intervenes when network control fails. Field experiment results show a maximum lateral error under 0.22 m with an average execution time under 20 ms, ensuring safe driving even in complete network control failure.