Employing Convolutional Neural Network for Real-time Detection and Localization of Physical Contact in a Redundant Manipulator Body

In the developing field of human-robot collaboration, robots increasingly share their workspace with humans, which becomes necessary to achieve seamless and safe interactions for manual guidance of robots in physical cooperation scenarios. The present study proposes a new method based on a convolutional neural network (CNN) for safe human-robot cooperation in the presence of intentional physical interactions, which is able to detect the occurrence and localization of contact on the robot arm in real-time. In this approach, contact detection and its localization are done only using the online external torque vector obtained by the residual-based observer of the robot, which eliminates any need to use additional force and vision sensors. Unlike traditional methods that depend on predefined thresholds or model-based techniques, our presented architecture with one-dimensional CNN can automatically learn the patterns from collected raw torque data and, by analyzing them, can properly predict the contact position. The performance of the proposed approach during manipulation has been validated through multiple online experiments in static and dynamic states on a KUKA LBR iiwa robot.