Human-Robot Contact Detection in Assembly Tasks

Abstract

Industry 5.0 complements the existing Industry 4.0 paradigm by, amongst others, focusing on flexible and human-centered manufacturing. However, there are some challenges when humans and robots share a workspace, such as human safety and limited robot perception. Addressing the above-mentioned challenges, we present in this paper a modular contact detection architecture for physical human-robot collaborations, especially for assembly tasks. This architecture consists of two sub-modules: a torque regressor using Gaussian Process Regressor (GPR) and a contact classifier using a Convolutional Neural Network (CNN). In fact, the GPR calculates deviations from expected joint torque values and the CNN detects object/human contacts with the robot, based on the estimated torque values. The result of the real-time implementation on an actual robot shows that the model can achieve a balanced accuracy of over 99% when the robot speed is over 27% and below 45% of its maximum speed even though the dataset was gathered with a speed of only 25%. This indicates the model's generalization capability to higher robot speeds. In addition, the generalization capability of the modular architecture for new movements was tested. The results show, that retraining the GPR can handle the data distribution shift and consequently, the contact classifier is not required to be retrained. Solving the impact of data distribution shift on the contact classifier makes the proposed modular architecture superior to the current state-of-the-art models.