A bi-level cooperative operation approach for AGV based automated valet parking

Abstract

In urban areas, the number of cars has increased significantly in recent years, resulting in frequent traffic congestion in parking lots. Automated valet parking (AVP) system based on automated guided vehicles (AGVs) can relieve human from parking and improve efficiency to a certain extent due to their fully automatic control and operation. However, with the expansion of the scale of the whole parking lot, the current AGV based AVP system is facing the disadvantage of long-time queue congestion and even deadlock. In this paper, we systematically consider the traffic congestion faced by the AGV based AVP system and introduce a bi-level cooperative operation approach. The global cooperative parking space allocation is considered in the upper-level, and the cooperative driving of multiple AGVs in the conflict zone is resolved in the lower-level. The upper-level problem is formulated as a Markov decision process, and a global cooperative allocation method is obtained by using deep reinforcement learning (DRL). In the lower-level, with the modified planning based cooperative driving method, multiple AGVs can drive efficiently without collision and deadlock in the conflict zone. Experiment results show that the proposed cooperative operation approach can significantly alleviate the congestion problem in the AGV based parking lot and improve the AVP system’s efficiency.