Modelo neuromimético para localização de estímulos táteis em pele eletrônica

Abstract

Tactile sensations are fundamental for social interaction and in our interaction with the world around us, as well as being crucial for bodily integrity. Through touch, we identify, locate, and react to external stimuli—functionalities that are desirable in artificial systems such as bionic prosthetics and collaborative robots. The development of biomimetic tactile systems, including tactile sensors, electronic skins (e-skins), and bioinspired models, would enable the development of new methods of tactile feedback in prosthetics and robotics. However, integrating biomimetic intelligence into these systems and creating flexible electronic skins with large areas and high sensory density remains challenging. A promising solution involves Fiber Bragg Grating (FBG) sensors, multiplexed in a single optical fiber embedded in soft polymeric substrates. These sensors mimic type II mechanoreceptors due to their large receptive fields. However, the integration of biomimetic models into such devices still faces important challenges, possibly due to the incomplete understanding of the biological mechanisms associated with the processing of tactile information. It is postulated that the temporal-spatial conversion of tactile stimuli from the upper limbs occurs in the cuneate nucleus. This spatiotemporal integration would then be crucial for our ability to perceive several aspects of tactile stimuli, such as their location on the skin. Thus, the objective of this thesis is to develop and evaluate a neuromimetic model of the cuneate nucleus to infer the location of mechanical indentations applied to an e-skin with FBG sensors. The model should combine the two main paradigms of the somatosensory system: the existence of overlapping receptive fields with a functional organization and the neuroplasticity dependent on the firing time of mechanoreceptors. For implementation and validation of the proposed model, an electronic skin with 21 FBG sensors was used, which was subjected to an automatic indentation protocol with 1846 target points, and the signal acquired from each FBG sensor was used as input for the neuromimetic model. The proposed model is divided into two bioinspired neural layers. The first layer models neural activity of slowly and rapidly adapting type II primary afferents (APs), in which the signal from each FBG sensor was multiplexed into six AP models, resulting in 126 APs. The second neural layer models the intracellular dynamics of the cuneate nucleus together with the modeling of receptive fields of functionally organized cuneate neurons (CNs) - this layer is composed of 1036 CNs and their inhibitory interneurons. The output triggers (action potentials) of the NCs were used to estimate the location of the stimuli applied to the e-skin. The results obtained by validating the model demonstrated that it is capable of learning in an unsupervised manner to locate stimuli applied to the electronic skin under different conditions of force and stimulus locations, with a median location prediction error of less than 10 mm for the sensorized region of the e-skin. These results demonstrate the generalization capacity of the proposed neuromimetic model and point to its potential use as a neurocomputational tool for research on the mechanisms of somatosensory processing, in addition to opening avenues for research on tactile feedback for prosthetics and robotic systems, based on the incorporation of intelligence.