Design and Validation of a Lightweight 1D CNN for Affective Touch Classification in Soft Plush Companions

Technical Summary: Design and Validation of a Lightweight 1D CNN for Affective Touch Classification in Soft Plush Companions

Problem Statement
Soft, sensorized interactive companions offer a safe and emotionally intuitive interface for socially assistive technologies (SAR), particularly for individuals with Autism Spectrum Disorder (ASD). However, the deformability of soft materials and the complexity of multichannel tactile sensing complicate the robust interpretation of human affect. Previous iterations of such systems relied on heuristic, threshold-based algorithms to classify touches as "positive" or "negative." These heuristics proved inconsistent, often misclassifying complex gestures (e.g., multi-sensor contacts like hugs) and failing to detect subtle social touches. Furthermore, reliance on external computation or cloud infrastructure limits portability, privacy, and the emotional authenticity of the interaction. The core challenge is to develop a computationally efficient, embedded deep learning model capable of real-time affective touch recognition on low-power microcontrollers without external connectivity.

Methodology
The study presents an end-to-end framework for developing and validating a compact deep learning system for a soft plush toy equipped with 13-channel capacitive sensors and a 3-axis accelerometer.

• Data Collection: A diverse, FAIR-compliant dataset was collected from 25 participants (children, teenagers, and adults). It comprises 1,326 labeled gesture sequences covering 17 distinct gestures, including positive (stroking, scratching), neutral (holding), and negative (hitting, pulling) interactions. Data was recorded under two conditions: the toy resting on a table and being held in hands.
• Preprocessing: The workflow involved anonymization, quality control (outlier detection), and interactive truncation to remove latency. Data was segmented using a sliding window of 250 samples (2.5 seconds at 100 Hz) with a 50-sample step. Short sequences were looped rather than zero-padded to preserve physical signal properties, expanding the dataset to approximately 5,600 training windows.
• Model Architecture: The study systematically explored 13 distinct One-Dimensional Convolutional Neural Network (1D CNN) architectures, varying the number of layers (1–4), filter counts, and the use of dilated convolutions to increase the receptive field without increasing parameter count. The models were optimized for embedded deployment on an ESP32-WROVER-E microcontroller.
• Training and Validation: 468 distinct models were trained using the Adam optimizer. Evaluation included random window-level splitting for initial architecture selection and Leave-One-Subject-Out Cross-Validation (LOSO-CV) to assess generalization to unseen users. Traditional machine learning baselines (Random Forest, SVM, kNN) using handcrafted features were also evaluated for comparison.
• Real-Time Simulation: Selected models were validated in a PC-based real-time simulator connected via Bluetooth to the physical toy, allowing for qualitative assessment of latency and classification consistency against the previous heuristic system.

Key Contributions

1. FAIR-Compliant Dataset: The public release of a diverse dataset of 1,326 labeled affective touch sequences from 25 participants, addressing the scarcity of open resources in affective touch recognition.
2. Open-Source Framework: A complete, reproducible MATLAB-based pipeline covering data acquisition, preprocessing, systematic CNN architecture exploration, and real-time simulation.
3. Optimized Lightweight Model: Identification of a compact dilated 1D CNN architecture (specifically 14d1_39d2_41d4 with 13.2k parameters) that achieves high accuracy while fitting within the memory and computational constraints of an ESP32 microcontroller.
4. Hybrid Inference Strategy: The proposal of a two-stage pipeline combining instantaneous heuristic filtering for high-force negative interactions with CNN-based classification for nuanced and neutral gestures.

Results

• Model Performance: The selected 1D CNN architecture achieved a test accuracy of 75% and a mean LOSO-CV accuracy of 85% (±0.072). The model demonstrated robust generalization across unseen subjects, with high F1 scores for positive and neutral gestures (e.g., Head stroke, Neck scratch).
• Comparison with Baselines: While traditional ML models (e.g., Random Forest) achieved higher accuracy (up to 95.9%) on random 5-fold cross-validation, they required significantly larger model footprints (up to 10 MB) and relied on handcrafted features. The 1D CNN operates directly on raw sensor data and fits within the embedded memory budget.
• Real-Time Validation: In PC-based simulations, the CNN successfully detected subtle gestures (e.g., ear scratching, tail tapping) that the previous heuristic system missed (6% activation vs. 100% for CNN). However, the heuristic system reacted faster (0.1–0.2 s) to high-energy negative gestures (e.g., tail hitting) compared to the CNN (2.0–2.5 s).
• Embedded Feasibility: Theoretical analysis indicates that a quantized version of the model requires approximately 3.2 MMAC per inference window. This translates to an estimated inference time of ~23 ms on the ESP32, supporting real-time operation at 20 Hz.

Significance and Claims
The paper claims that emotionally meaningful, privacy-preserving touch interpretation is computationally feasible for direct embedding within soft therapeutic companions. By demonstrating that a lightweight 1D CNN can resolve subtle social touches better than heuristic methods while maintaining low latency and low power consumption, the study establishes a scalable foundation for emotionally intelligent, soft-bodied smart companions. The work bridges the gap between high-level affective computing and resource-constrained edge hardware, moving SAR research from rigid, externally computed systems toward autonomous, soft, and emotionally responsive devices. The authors note that while the model design and validation are complete, the physical integration and field testing in therapeutic sessions are designated for a forthcoming study.