Real-time Obstacles Avoidance for Crawler Crane based on DQN

Abstract

Safety always is a crucial consideration for crawler cranes running with dozens of tons of load in its work site. However, around its work site, besides static obstacles, there are many dynamic obstacles, such as workers and engineering vehicles. These obstacles are potential risk for crane in congested work site. Therefore, risk probability and loss can be reduced in project of lifting engineering by dynamic lift-path planning system. However, presently, most proposed methods pay attention to search path in totally known environment. In this paper, we present a dynamic obstacles avoidance planner for crawler cranes in no-walk scenarios in partially known environment. This planner considers the lifted module as a 3DOF convex robot with discrete rotational and translational motions. First, we improved the structure of neural network of Deep Q-network by applying Resnet block for achieving real-time path planning for crawler crane. Thereafter, we also improved Artificial Potential Field Method (APFM) and apply it to search path for lifted module. And for shortening training time, we also apply the data generated by APFM to pre-train our neural network. Finally, we verified and compared the performance of APFM and DQN by simulation. The result of simulation can demonstrate the effectiveness of the presented planner.