Lightweight CNN Classifies Affective Touch in Soft Companions

Soft, sensorized companions offer a promising interface for socially assistive technologies, but interpreting human affective touch from their deformable surfaces and multichannel tactile sensors is challenging. This research introduces a complete open-source MATLAB-based framework for developing and validating compact deep learning models specifically for affective touch recognition in these interactive companions. A significant contribution is a publicly available dataset of 1326 labeled gesture sequences from 25 participants, including children, teenagers, and adults, providing a valuable resource for future research. Through extensive architectural and hyperparameter exploration, the study identified compact dilated one-dimensional convolutional neural networks (1D CNNs) as the most effective solution. A model with just 13.2k parameters achieved 75% test accuracy and 85% mean leave-one-subject-out cross-validation accuracy. Theoretical analysis suggests that a quantized version of this model could operate in real-time on target microcontrollers. Real-time simulations with a physical toy demonstrated the CNN's ability to resolve subtle social touches that previous heuristic systems missed, while high-force negative interactions were reliably captured by simpler threshold logic. The proposed embedded deployment strategy combines instantaneous heuristic filtering with CNN-based nuanced gesture classification, proving that emotionally meaningful, privacy-preserving touch interpretation is computationally feasible for direct integration into soft therapeutic companions.