강화학습을 이용한 자율주행 트랙터의 가변형 전방주시거리 튜닝 알고리즘 개발

Abstract

Robust path tracking is important for an autonomous tractor to accurately track a path, including straight and curved paths, with headland turning in farmlands. To improve the performance of a path-tracking method developed on the basis of fixed look-ahead distances, this paper presents a look-ahead distance tuning algorithm based on reinforcement learning (RL). The RL architecture is designed to provide variable look-ahead distances in real time based on navigation conditions caused by changes in the lateral deviations and heading errors. The RL agent performed self-learning in the simulation while following C-type turning-based paths. For more effective RL, a simulation model was built by applying the actual responses of a tractor steering system measured using a data acquisition system. An evaluation test was conducted in terms of the lateral deviation in response to the reference path. The proposed method improved the path-tracking accuracy of the lateral deviation by a root mean-squared error of about 62% compared to a path-tracking controller with fixed look-ahead distances.