Humanoid Robot Classification via Xception Deep Neural Network

Abstract

Humanoid robots have emerged as a key research field in recent years, driven by their potential in diverse applications, including industry, entertainment, assistance, logistics, and others. In this work, we propose a convolutional neural network architecture for the classification of state-of-the-art humanoid robots using the Xception Convolutional Neural Network (CNN). A comprehensive dataset comprising 5,399 RGB images across 31 different humanoid robot classes was curated from online sources, including official websites and public multimedia resources. To enhance model performance, data augmentation techniques were applied, including random rotations, reflections, translations, scaling, and shearing, with all images resized to a uniform pixel resolution. The dataset was divided into 70% training, 15% validation, and 15% test subsets for model evaluation of overfitting and underfitting. The Xception model achieved high performance on the test set, with average metrics across classes of 99.72% specificity, 79.00% precision, 71.46% recall, 75.34% F1-score, 0.759 MCC, and 0.608 Jaccard index. These results highlight the model’s high accuracy and robustness in real-time humanoid robot classification, demonstrating its potential for applications in dynamic, real-world environments, such as in multi-humanoid collaborative environments.