Tuning of Model Predictive Control for Loading Motion of Automated Excavator Using Deep Reinforcement Learning

This study deals with the control problems for automating the operation of an excavator loading soil onto the back of a dump truck. In the loading operation, the bucket should not touch the dump truck and should spill as little soil in the bucket as possible. We have been studying how to apply Model Predictive Control (MPC) to this problem to achieve ideal loading operation. When trying to achieve the desired operation using MPC, it is extremely important to tune the weights of the objective function appropriately. However, since this control problem may depend on the situations, that is, initial posture of the excavator and the position of the truck, optimization for specific conditions would not be desirable. Therefore, we constructed a method to generate suitable weight parameters according to the loading situation using reinforcement learning. The effectiveness of the proposed method was verified by numerical simulations.