Leveraging human expert knowledge to automate forklift truck driving

Abstract

This work explores the challenges of fully automating in-house goods transport in environments where industrial trucks like forklift trucks remain necessary due to undefined load carrier positions and shapes. Imitation Learning (IL) is identified as a promising solution for vehicle control in repetitive tasks, yet its application in intralogistics is challenging by the dynamic complexity of industrial trucks and the large dimensional space involved. A Robot Operating System 2 (ROS2) framework is introduced, enabling the acquisition of driving data from both simulation environments and real-world demonstrators. The study also presents a network architecture combining a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) network, facilitating end-to-end learning from spatial and temporal image data. The framework's effectiveness is evaluated using a dataset of expert driving maneuvers to assess the generalization potential of the IL-trained network in vehicle control in different scenarios. The research aims to demonstrate the utility of the proposed framework for data acquisition and validate IL as a control approach for industrial trucks that require generalization