Neural Networks for Tactile Perception.

dc.contributor.authorPati, Y.C.en_US
dc.contributor.authorFriedman, Daniel E.en_US
dc.contributor.authorKrishnaprasad, Perinkulam S.en_US
dc.contributor.authorYao, C.T.en_US
dc.contributor.authorPeckar, M.C.en_US
dc.contributor.authorYang, R.en_US
dc.contributor.authorMarrian, C.R.K.en_US
dc.description.abstractIntegrated tactile sensors appear to be essential for dextrous control of multifingered robotic hands. Such sensors would feature (1) compliant contact surfaces, (2) high resolution surface stress transduction, (3) local signal conditioning, and (4) local computation to recover contact surface stress. The last-mentioned item pertains to the basic inverse problem of tactile perception and the real time solution of this inverse problem is our primary concern. We think that good solutions to this problem (i.e., algorithms + implementations) will be needed for realizing dextrous hand control via tactile serving. In this paper we describe a processor chip designed to solve the mathematical inversion problem utilizing neural network principles. Simulations indicate that this chip can function in the presence of large amounts of electrical noise. In addition the effect of processing induced variability in sensor response can also be minimized using the maximum entropy estimate method described below. The tactile sensor design we refer to is the one reported in [1]. This particular design is based on piezo- resistive transduction via an array of diffuse resistors in silicon. Surface load on a compliant layer is transformed into resistance changes proportional to biaxial strains. Initial testing of the sensor has yielded repeatable, linear characteristics. The signal conditioning chip which acts as an interface between the sensor array and subsequent processor chips has also been fabricated. The neural network chip described in this paper has been simulated at the system level. The simulation results for this network based on a particular linear elastic model (described in section 2) of the compliant contact layer. We consider in the simulations some of the errors introduced by process variability in VLSI implementation. The simulations carried out using SIMNON a general purpose nonlinear simulation package developed at Lund Institute of Technology, Sweden (kindly provided us by Professor Astrom), are described in section 4.en_US
dc.format.extent663561 bytes
dc.relation.ispartofseriesISR; TR 1987-123en_US
dc.titleNeural Networks for Tactile Perception.en_US
dc.typeTechnical Reporten_US
Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
648.01 KB
Adobe Portable Document Format