Linear Optimal Transport for Domain Adaptation in Neuromorphic Tactile Texture Classification

This paper introduces a neuromorphic tactile sensing system based on piezoelectric polymers for texture classification. Raw tactile signals were first converted into spike trains, from which spike-based features were extracted to train a multilayer perceptron (MLP). To improve generalization and mitigate the so-called batch effect, an optimal linear transportation was applied along with a leave-one-subject-out evaluation strategy. Experimental results show that the model trained on the transported data achieved a performance improvement of approximately 90% in accuracy compared to the original data. The proposed system demonstrates strong potential for subject-independent tactile perception. This dataset contains raw tactile signals acquired using a flexible piezoelectric sensing system during texture exploration experiments involving multiple participants. The sensing system consists of an array of eight P(VDF-TrFE) piezoelectric sensing units screen-printed on a flexible substrate and integrated into a soft biomimetic fingertip. The eight tactile channels were acquired simultaneously at a sampling frequency of 2 kHz. The dataset was collected from three participants exploring 12 naturalistic textures characterized by different surface properties, including soft, smooth, and rough materials. Each participant wore the sensorized fingertip on their index finger and explored the textures by sliding in different directions, without controlling the applied force or exploration velocity. The dataset comprises 36 multichannel recordings, corresponding to one recording per texture for each participant. The recordings are organized by participant and texture class and provided as raw time-series signals in tab-separated LabVIEW Measurement (.lvm) files. Each recording contains eight sensing channels, preserving the temporal information acquired during the texture exploration process. The dataset was used to investigate neuromorphic approaches for tactile texture classification and domain adaptation across different participants. In the associated study, raw tactile signals were converted into spike trains using the Izhikevich neuron model, and spike-based features were extracted to train a multilayer perceptron (MLP). A linear optimal transport approach was applied to mitigate subject-dependent variability in the feature distributions and improve classification performance under a leave-one-subject-out evaluation strategy. The dataset provides a resource for investigating neuromorphic tactile encoding, texture recognition, cross-subject generalization, and domain adaptation in piezoelectric sensing systems. Only the original tactile recordings are provided; spike trains, extracted features, trained models, and processing code are not included. The dataset is associated with the following scientific publication, which provides a detailed description of the sensing system, experimental protocol, neuromorphic encoding methodology, and domain adaptation framework: L. Hamadi, M. Yaacoub, H. Haj Ali, M. Valle, and Y. Abbass, "Linear Optimal Transport for Domain Adaptation in Neuromorphic Tactile Texture Classification," in Applications in Electronics Pervading Industry, Environment and Society (ApplePies 2025), Lecture Notes in Electrical Engineering, vol. 1553, pp. 261–266, Springer, Cham, 2026. DOI: https://doi.org/10.1007/978-3-032-17174-0_37 Users of this dataset are kindly requested to cite the above publication when using the data in their research.