Development of a Face Recognition System Based on a Convolutional Neural Network on a Humanoid Robot

This paper presents the development of a face recognition system based on a Convolutional Neural Network (CNN) for deployment on the Unitree G1-EDU Humanoid Robot. By integrating the robot's built-in Automatic Speech Recognition (ASR), audio, and motion capabilities, the system enables the robot to deliver context-aware audio responses and physical reactions (e.g., waving or shaking hands) based on face recognition results when a person is detected in front of it. We evaluated the system's performance using various CNN models across multiple face recognition datasets and conducted real-time tests with students in our lab. The estimated mean accuracy (ACC) was used as the assessment metric. The offline evaluations achieved an average accuracy of over 95 % across three benchmark datasets using the best model, while the real-time tests reached approximately 91 % accuracy. These results demonstrate the feasibility of deploying an efficient CNN-based face recognition system on a humanoid robot with limited computational resources and highlight the effectiveness of lightweight CNN models for realworld face recognition tasks.