No abstract

In this paper we present a performance analysis of a perceptual architecture for industrial robots based on tactile sensing. We have developed a human-robot interface based on sensing devices which can be fixed on the robot body or the gripper. The devices are cylindershaped handles covered with tactile sensors, which can be intuitively grasped by an operator. Tactile data are fed into a neural network in order to recognize human touch during grasp, thus providing an enabling command for the control system. We provide an inference time comparison between two computing architectures: a desktop workstation with an high-performance GPU and an embedded solution based on the NVIDIA Jetson Nano board. We also compare inference time obtained from three instances of the neural network, compiled with three different engines: Keras, TensorRT floating-point 16 and TensorRT floating-point 32, showing that numerically optimized models allow to perform inference also on the embedded board without violating timing constraints imposed by data acquisition. Moreover, we assess the robustness of touch recognition when the user is wearing work gloves, showing that the difference from the bare-hand case is negligible.

No abstract

No abstract