Tactile sensors represent an intuitive and efficient interface for physical Human-Robot Interaction (pHRI). To ensure the safety of the human during industrial collaborative tasks we mounted four cylindrical-shaped handles, covered with the CySkin technology, over the gripper of an industrial manipulator, through which the operator can express his intention to physically interact. Tactile data are fed into a neural network that recognizes human touch during grasp, thus providing an enabling

more »

... and for the control system. In this paper we present a performance analysis of the perceptual architecture based on distributed tactile sensors for Human-Robot Collaboration (HRC). The inference time comparison performed in this activity considers two computing architectures: a desktop workstation with a high-performance GPU and an embedded solution based on the NVIDIA Jetson Nano board. The inference time performance analysis also considers three different neural network optimization engines: Keras, TensorRT floating-point 16 and TensorRT floating-point 32. The results show that numerically optimized models allow to perform inference within the timing constraints even on the embedded architecture. An inference robustness analysis is also performed to verity if voluntary touch recognition fails when the user wears work gloves, showing negligible differences from bare-hand grasps.