RAC-Touch: Reservoir Computing and Active Touch Based New Architecture for Larger, Faster and Robust E-skins

Description

The sense of touch is essential for us to thrive in vision-denied situations (e.g., replacing a light bulb). Akin to humans, touch is a critical sense in robots, especially to enable them to work alongside humans in unstructured environments and interact naturally with them. Such artificial skin technology, termed electronic or e-skin, is expected to allow for social robots in senior facilities, teleoperated surgery, sensory feedback in prosthetic devices, online shopping, etc. However, despite this apparent need to digitize the sense of touch, it has proved much more complex than digitizing vision or auditory senses, due to the following conflicting requirements: (1) high number of touch pixels or taxels (2) fast sampling rate from the taxel array (3) less wires to make a robust system (4) low data rate to enable quick processing and feedback in closed-loop system design. This proposal aims to take inspiration from biology and create a neuromorphic e-skin with three main novelties: (1) Instead of reading out the raw values of all sensors, they are modulated using feedback voltages and combined onto a single wire. Recent advances in machine learning (reservoir computing) show how such a combined signature can be used to classify the information in the original sensor array values. (2) Human senses are developed jointly with motor actions defined by behavior through an action-perception loop. We propose to use feedback from the sensed values and task goal to decide further the following sequence of feedback voltages and arm movements to be done to maximize information transfer to the e-skin thus changing the sensing to “active” touch. (3) To reduce data rates, the raw signal from the e-skin is analyzed for significant changes from the baseline, and scanning is performed only for those cases. This event-driven approach reduces the average data rate significantly. By combining Reservoir computing with ACtive touch, we expect our proposed RAC-touch e-skin to reduce wires by ~50%, keep average latency < 1ms, reduce average data rate by >10X while retaining classification accuracy close to that of the original taxel array by exploiting active sensing through feedback. The results will be demonstrated using a robot arm as well as a human worn tactile glove using our e-skin.