level: research
soft, sensorized companions can provide safe and intuitive interaction for socially assistive technology, but their squishy bodies and many tactile sensors make it hard to reliably read human emotion from touch. this work introduces an open-source matlab framework for building and testing small deep learning models that classify affective touch on soft interactive companions. the main contribution is a fair-compliant public dataset with 1326 labeled gesture sequences, collected from 25 people including children, teenagers, and adults, giving future research a reusable benchmark.
the authors ran a large hyperparameter search across 468 convolutional neural network designs. they found that compact dilated one-dimensional cnns work best for this task. the winning model uses only 13.2k parameters, making it suitable for embedded deployment on low-power hardware inside a plush toy. the study details how different architectural choices affect accuracy and efficiency, providing practical guidance for similar sensor-rich soft robot applications.
the framework and dataset lower the barrier for developing emotion-aware soft robots. by sharing code and data openly, the work enables other researchers to replicate results and build on the findings without starting from scratch. the lightweight model shows that high recognition performance does not require large, computationally heavy networks, which is important for battery-powered, interactive devices that must respond in real time.
why it matters: it provides a ready-to-use dataset and a tiny neural network design that can run on cheap hardware inside soft toys, helping developers build emotionally responsive companions without heavy computing resources.