The human sense of touch plays a critical role in tactile perception and dexterity so it seems logical to conclude that robotic systems seeking these capabilities will require tactile sensors. We have developed such an artificial tactile sensor (the BioTac) that has similar mechanical properties and sensory capabilities as the human fingertip. This device was used to objectively quantify the tactile properties of everyday surfaces, even outperforming human subjects in AB discrimination tasks of challenging surfaces. Achieving this level of performance was done in a very biomimetic process using: tactile sensors that behaved similarly to human fingertips, exploratory movements similar to those humans make when exploring objects, signal processing inspired by language and neuroscience, and a newly developed decision-making algorithm we’ve called Bayesian exploration. Early findings indicated that tactile classification introduced new challenges that were not present in image and audio classification, namely that as touch was an active sense the selection of exploratory movement has a tremendous effect on what you feel. This naturally brought the question of which exploratory movements were optimal for eliciting the type of sensory data for best classifying surfaces. When humans perform this task of identifying objects by touch they do not always perform a fixed set or sequence of movements, instead, they adaptively make movements that depend on their evolving hypothesis of what that object is and the information being sensed. To replicate this process, an algorithm called Bayesian exploration was developed to use previous experience to determine which of many candidate movements would yield data to best disambiguate between the current hypotheses of what that surface may be, as quantified by a probability vector. After performing this exploratory movement Bayesian inference would be used to update that probability vector and the Bayesian exploration algorithm would then adaptively determine the next optimal movement until a classification was made. It was found that with these methods a surface could be quickly identified from 500 candidate surfaces in on average 2.5 exploratory movements with greater than 99.5% accuracy, far exceeding human capabilities. While the potential for these tactile characterization and classification tools could provide great benefit to a new generation of medical tactile imaging, it was also realized that the decision-making algorithm to achieve this performance shares a great deal of similarity with the process of diagnostic medicine and that perhaps the characteristic of Bayesian exploration to effectively handle noisy information sources could also provide new opportunities for decision-making support based on electronic health record data.


Author: Jeremy Fishel

Status: Work In Progress

Funding Acknowledgment: This material is based upon work supported by the National Science Foundation under Grant Nos. 1345335 and 1534524.