Lightweight 1D CNN reads affective touch from plush toys at 85% accuracy

A team of researchers led by Aleksandrs Vališevskis has published a framework for affective touch classification in soft, sensorized plush companions. The system uses a lightweight 1D convolutional neural network (CNN) with only 13.2k parameters, designed for embedded deployment on microcontrollers. The team systematically explored 468 CNN architectures and selected a dilated 1D CNN as the most effective. Using a FAIR-compliant dataset of 1,326 labeled gesture sequences from 25 participants spanning children, teenagers, and adults, the model achieved 75% test accuracy and 85% mean leave-one-subject-out cross-validation accuracy. The quantized version requires just 3.2 multiply-accumulate operations per window, enabling 20 Hz real-time performance on a target microcontroller.

In real-time tests with a physical soft companion streaming sensor data, the CNN successfully resolved subtle social touches that the previous heuristic system missed. However, high-force negative interactions were better captured by simple threshold-based logic. The final design is a hybrid inference pipeline: instantaneous heuristic filtering for obvious high-force events, followed by CNN-based classification for nuanced gestures. The authors propose this as the embedded deployment strategy for soft therapeutic companions, offering privacy-preserving, emotionally meaningful touch interpretation that is computationally feasible for direct integration. Hardware integration details are promised in a forthcoming study.