A Tactile Sensor-Based Architecture for Collaborative Assembly Tasks with Heavy-Duty Robots

Abstract

In this article we present a sensing device to enable safe and reactive human-robot interaction for collaborative tasks involving industrial heavy-duty robots. We have designed and implemented an handle covered with artificial skin intended to provide commands in response to operator's touch. The tactile processing system is capable to recognize human voluntary interactions from tactile images by exploiting a pre-trained convolutional neural network. We show the accuracy results of a fine-tuning training of the network done using new data acquired from the handle. We also provide a performance analysis of the model inference on a Nvidia Jetson Nano board, confronting the inference time from two different optimization models, fp16 and fp32 and showing that fp16 better performs in terms of temporal predictability.