Biomimetic tactile sensor and spike train processing for surface roughness discrimination and active exploration
Date of Issue2017
School of Mechanical and Aerospace Engineering
The sense of touch plays a critical role in enabling human beings to interact with the surrounding environments. As robots move from laboratories to domestic environments, they are expected to be endowed with a similar tactile ability to perform complicated tasks such as manipulating objects with arbitrary unknown shapes. This thesis investigates the approaches to improve the tactile capabilities of future robotic systems in two aspects, i.e., biomimetic tactile sensing for surface roughness discrimination and active tactile exploration for object shape reconstruction. Surface texture discrimination using artificial tactile sensors has attracted increasing attentions in the past decade. However, as a major component of texture, surface roughness has rarely been explored. We present two signal processing methods for tactile surface roughness discrimination based on a biomimetic fingertip. One method focuses on classical machine learning algorithms. Specifically, various extracted features and classification algorithms including support vector machines (SVM) and k-nearest neighbors (kNN) are examined. The other method is based on spike-like signal analysis. The analog tactile signals generated from polyvinylidene difluoride (PVDF) films are fed as input to the Izhikevich neurons in order to obtain spike trains. Two distinct decoding schemes based on k-nearest neighbors (kNN) in both spike feature space and spike train space are used for surface roughness discrimination. We thoroughly examine the different spike train distance based kNN (STD-kNN) algorithms for decoding spike trains. Eight standard rough surfaces with different surface roughness are explored. We find that the highest classification accuracy of (82.6 ± 10.8) % can be achieved using solely one PVDF film sensor with kNN (k = 9) classifier and the standard deviation feature. For the soft neuromorphic approach, the highest classification accuracy of (77.6 ± 13.7) % can be achieved with kNN (k = 11) classifier and the Victor-Purpura distance (q = 0.024 ms-1)). In the tactile exploration task, we present an active touch strategy to efficiently reduce the surface geometry uncertainty by leveraging a probabilistic representation of object surface. In particular, we model the object surface using a Gaussian process and use the associated uncertainty information to efficiently determine the next point to explore. We validate the resulting method for tactile object surface modeling using a real robot to reconstruct multiple, complex object surfaces.