On Edge Pinch Force Classification Using IMU Sensors and Convolutional Neural Networks

This paper introduces an embedded system designed for the real-time classification of force levels using Inertial Measurement Unit (IMU) sensors and Convolutional Neural Networks (CNNs). The system's objective is to enhance proprioceptive capabilities in robotics, enabling precise object manipulation by accurately detecting and interpreting forces applied to joints. Two CNN architectures are proposed and evaluated: one that processes combined accelerometer and gyroscope data, and another that processes these data separately before combining the extracted features. Both approaches were deployed on a low-power STM32F401 microcontroller, demonstrating effective real-time performance. The first approach achieved a classification accuracy of 88.6%, while the second achieved 87.9%, with both approaches consuming approximately 21 mJ per inference.