Optical tactile force sensing with a soft transparent sensory skin using laser time-of-flight measurement and nonlinear modelingZore, Aleš (Avtor) Babić, Marina (Avtor) Munih, Marko (Avtor) Kamnik, Roman (Avtor) applied contact force predictionLIDARlight-guideneural networkNNsoft sensor skintactile sensinglight detection and rangingtime of flightToFThis work presents a machine learning-based framework for predicting applied contact force from the optical response of a time-of-flight (ToF) sensor embedded in a soft, optically transparent tactile sensing medium, SoftSkin. The SoftSkin medium acts as a light-guiding structure, enabling light propagation via total internal reflection. External pressure deforms the light-guide material, resulting in measurable changes in the optical response. A robotic test setup equipped with a force–torque sensor enabled controlled and repeatable experiments. Training datasets were collected at three indentation velocities and six applied force levels. Three multilayer perceptron (MLP) architectures were developed and evaluated. The single-input method used only photon counts, whereas the dual-input method incorporated temporal ambient-light conditions. A third approach, the derivative-enhanced dual-input method, additionally integrates temporal gradients of the photon response. For each MLP model, a Random Forest (RF) residual corrector was applied to reduce prediction errors. Performance was evaluated on an independent test set comprising unseen force magnitudes at three indentation velocities. All MLP–RF models achieved a mean absolute prediction error below 1 N, with the derivative-enhanced dual-input model demonstrating the best performance. The results confirm that ToF-based optical tactile sensing, combined with lightweight neural network architectures and residual learning, enables accurate reconstruction of dynamic force profiles despite sensor noise and the limited frame rate of the ToF sensor.20262026-09-15 08:57:53Članek v reviji187829sl