Predictive tactile coding in a memristive neuromorphic tactile system

Abstract Flexible tactile electronics have progressed rapidly in wearable electronics, electronic skin, tactile sensing materials and neuromorphic sensory devices, yet most artificial touch systems still interpret contact only after the signal has already been measured. Such a reactive mode supports static recognition, but it is less suited to continuous interaction, where tactile sensing is embedded in the evolving contact process and must follow how surface morphology changes during scanning. Inspired by predictive processing in biological sensory systems, in which incoming signals are interpreted against an internally generated expectation of the next sensory state, we introduce a predictive tactile neuromorphic system that couples a 44×44 flexible tactile array to an 8×8 memristive array for one-step next-state estimation and mismatch-based tactile inference. We program the learned latent-state transition matrix into the memristive conductance matrix so that the current encoded tactile state can be projected in hardware to its predicted successor. In experiments, a flexible pressure sensor mounted on a robotic fingertip records the evolving contact fields during sliding. Each tactile frame is compressed into an 8-dimensional latent representation and propagated through a programmed memristive conductance matrix to generate the predicted next tactile state, which is then compared with the measured next state to produce a normalized mismatch score that quantifies the deviation between expected and observed tactile evolution. Across 1,800 scan sequences spanning smooth, coarse periodic, and fine periodic surface states with matched local violations, the system maintains low mismatch during regular tactile evolution and generates pronounced mismatch peaks when local continuity is broken. The pooled anomaly-discrimination performance reaches an ROC AUC of 0.992. The proposed predictive framework achieves a normal-region prediction error of 0.018 a.u., where normal-region refers to scan segments outside the disturbed zone, and maintains AUCs of 0.96, 0.91, and 0.87 under three speed/force perturbation settings, compared with 0.78, 0.70, and 0.66 for the no-temporal-context baseline. These results show that expectation-driven tactile inference can be embedded directly into the sensing-computing pathway and can support compact neuromorphic touch systems for robotic inspection, dexterous manipulation, and adaptive human-machine interfaces operating under continuously changing contact conditions.