Applications of Reinforcement Learning to Effectively and Safely Control a Gantry Crane

Abstract

When operating a gantry crane, a widely used machine in construction and transportation, an operator must carefully balance between delivering the load quickly and safely. However, the task of dampening the swing of the load, both during its transport and when it is at the target location, poses a difficult task for human operators and causes inefficiencies in this transportation process. PID controller systems, which is the standard in industrial applications, have had limited success in addressing this issue, facing difficulties controlling a nonlinear system or an inability to strictly enforce safety requirements. To address this issue, I test the viability of using reinforcement learning to train an agent to control a gantry crane in an efficient and safe manner. I demonstrate the success of agents trained with reinforcement learning to deliver a load to its destination in different environments and under various success conditions, even outperforming standard PID controllers. Additionally, I demonstrate the ability of these controller agents to learn novel strategies, suggesting their efficacy in learning in more complex environments or with additional constraints.