Discrimination-Based Perception for Robot Touch
Biomimetic tactile sensors often need a large amount of training to distinguish between a large number of different classes of stimuli. But when stimuli vary in one continuous property such as sharpness, it is possible to reduce training by using a discri
- PDF / 1,123,274 Bytes
- 5 Pages / 439.37 x 666.14 pts Page_size
- 13 Downloads / 200 Views
1
)
School of Physiology, Pharmacology and Neuroscience, University of Bristol, Bristol, UK [email protected] 2 School of Experimental Psychology, University of Bristol, Bristol, UK 3 Department of Engineering Mathematics, University of Bristol, Bristol, UK 4 Bristol Robotics Laboratory (BRL), University of Bristol, Bristol, UK
Abstract. Biomimetic tactile sensors often need a large amount of training to distinguish between a large number of different classes of stimuli. But when stimuli vary in one continuous property such as sharpness, it is possible to reduce training by using a discrimination approach rather than a classification approach. By presenting a biomimetic tactile sensing device, the TacTip, with a single exemplar of edge sharpness, the sensor was able to discriminate between unseen stimuli by comparing them to the trained exemplar. This technique reduces training time and may lead to more biologically relevant models of perceptual learning and discrimination. Keywords: Robotics · Psychophysics · Tactile sensing · Biomimetics
1
Introduction
Almost all animals rely on tactile perception in order to successfully move around and interact with their environment: humans primarily use their hands, rats depend signifi‐ cantly on their whiskers, and cockroaches rely on their antennae. Many of these various tactile systems that exist in nature are now being used as inspiration in robot touch, with a range of tactile sensors available based on biologically-inspired principles [1]. But as demands are placed on these sensors to distinguish ever smaller differences in stimulus properties, they often require a very large, cumbersome amount of training to accurately discriminate between very similar classes. To make the training process more efficient, here we conjecture that when stimuli vary continuously over just one property (e.g. size or curvature or angle), the sensor can be trained on a single exemplar and subsequently a new previously unencountered stimulus can be compared with the exemplar. In tasks where there might be dozens of different stimuli, this new approach could reduce training time significantly. Moreover, this is likely to be a more biomimetic method of training, as humans can easily perceive object properties on continuous scales and make relative judgements about stimulus properties (e.g. bigger or smaller). To test this conjecture we use the TacTip [2], a biologically-inspired optical tactile sensor designed to mimic responses to skin deformation in human fingertips, which has been shown to perform very well on a range of tactile perception and identification tasks [3–5]. Here we demonstrate that after training on just one exemplar of a stimulus
© Springer International Publishing Switzerland 2016 N.F. Lepora et al. (Eds.): Living Machines 2016, LNAI 9793, pp. 498–502, 2016. DOI: 10.1007/978-3-319-42417-0_53
Discrimination-Based Perception for Robot Touch
499
property (in this case, edge sharpness) the TacTip can generalise this property and distinguish between new, unseen stimuli that
Data Loading...
