Autonomous Load Carrier Approaching Based on Deep Reinforcement Learning with Compressed Visual Information

Abstract

In intralogistics, a large number of tasks are already fully automated. This holds true especially for tasks where strictly predefined positions and paths are specified and implementable. Challenges still exist when this is not the case and a more dynamic environment is present. One example of such a dynamic environment is the approaching and lifting of freely positioned pallet-like carriers with forklifts. In this work, we propose a method for approaching and picking up pallet-like carriers with a forklift based on data from an RGB camera. Unlike previous work, our method does not require an estimation of the pose of the load carrier. In order to control the forklift, we use a soft actor critical reinforcement learning agent. The required input consists of the bounding box of the load carrier in combination with the current speed and steering of the forklift. Our simulation experiments show that this compressed visual information is sufficient to successfully approach load carriers while reducing training time and network size. In a next step, we are going to apply the presented result on a real-world scenario and investigate its transferability.