Study on the Autonomous Walking of an Underground Definite Route LHD Machine Based on Reinforcement Learning

Abstract

The autonomous walking of an underground load-haul-dump (LHD) machine is a current research hotspot. The route of an underground LHD machine is generally definite, and most research is based on the logic of positioning-decision control. Based on a reinforcement learning algorithm, a new autonomous walking training algorithm, Traditional Control Based DQN (TCB-DQN), combining the methods of traditional reflective navigation and reinforcement learning deep q-networks (DQN), is proposed. Compared with the logic of location-decision control, TCB-DQN does not require accurate positioning, but only determines how to reach the endpoint by sensing the distance from the endpoint. Through experimental verification, after using the TCB-DQN algorithm for training in a simple tunnel, the LHD machine could achieve a walking effect similar to that of a human driver’s manual operation, while after training in a more complex tunnel, the TCB-DQN algorithm could reach the endpoint smoothly.