Researchers release open dataset and compact 1D CNN model for recognizing affective touch in soft robotic companions

Researchers released an open-source MATLAB framework and a FAIR-compliant dataset comprising 1,326 labeled gesture sequences collected from 25 participants spanning children, teenagers, and adults. The dataset is intended as a reusable resource for affective touch recognition research in soft interactive companions.

Through systematic exploration across 468 CNN models, the team identified compact dilated one-dimensional convolutional neural networks (1D CNNs) as the most effective architecture for the task. A 13.2k-parameter 1D CNN achieved 75% test accuracy and 85% mean leave-one-subject-out cross-validation accuracy on the dataset.

Theoretical analysis of quantized inference showed the model requires 3.2 million multiply-accumulate operations per window, which is compatible with 20 Hz real-time operation on the target microcontroller. A PC-based real-time simulation using streaming sensor data from a physical toy demonstrated that the CNN resolved subtle social touches that a prior heuristic system failed to detect, while high-force negative interactions were captured more reliably by simple threshold-based logic.

The authors propose a hybrid inference pipeline—combining instantaneous heuristic filtering with CNN-based nuanced gesture classification—as the embedded deployment strategy. They argue that emotionally meaningful, privacy-preserving touch interpretation is computationally feasible for direct embedding within soft therapeutic companions, with hardware integration to be addressed in a forthcoming study.