EffiAtt-Emotion: A Lightweight Emotion Recognition Model for Adaptive Human-Robot Interaction in Collaborative Robotics

Abstract

Effective and safe interaction in collaborative robotics hinges on the robot's ability to adapt its behavior to the human's cognitive and emotional state. This paper introduces EffiAtt-Emotion, a lightweight deep learning model for facial emotion recognition designed specifically for resource-constrained onboard robotic systems. With only 13.3 million parameters, the model demonstrates a high and balanced performance, achieving an accuracy of 80.6% and a macro F1-score of 0.806 for eight distinct emotion classes. We present a framework for integrating EffiAtt-Emotion into a collaborative robot's (cobot's) control loop, enabling real-time adaptive behavior. A use case on a collaborative assembly line is detailed, where the cobot adjusts its speed, provides assistance, or enters a safe stop mode in response to the human operator's recognized emotions, such as confusion, frustration, or fear. This approach directly addresses documented challenges in HRI, where human factors like stress and fatigue contribute to decreased performance and safety risks. By making the robot aware of the operator's state, our system aims to enhance task efficiency, ergonomics, and critically, to improve safety in line with the human-centric principles of Industry 5.0.