همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Enhancing Imitation Learning for Humanoid Robots Using Vision Transformers and Time Contrastive Networks

نویسندگان :

Amirmohsen Sharifi¹ Maziar Palhang²

1- Isfahan University of Technology 2- Isfahan University of Technology

کلمات کلیدی :

Imitation Learning،Vision Transformers،Time Contrastive Networks،Humanoid Robotics،Deep Reinforcement Learning،NAO Robot

چکیده :

Abstract— This paper describes a new method for skill imitation learning for humanoid robots using Vision Transformers and Time Contrastive Networks. For learning the temporal representations of robot gestures, we evaluate two leading visual encoding models, ViT B16 and EfficientNetV2S. The use of a triplet loss function with hard positive and negative mining ensures that the robot learns discriminative embeddings that capture the subtle temporal differences of consecutive video frames. The generated embeddings provide dense reward signals for deep reinforcement learning, enabling the NAO humanoid robot to learn gesture imitation tasks in the Webots simulation platform. The experimental results indicate that, of the tested configurations, the optimized ViTB16 architecture consistently surpassed the fine-tuned EfficientNetV2S model. The use of three fold cross validation with GroupKFold stratification serves to strengthen our results. The embeddings produced by ViT B16 allow for the formation of well-separated and structured latent representations as evidenced by PCA, tSNE, and UMAP visualizations, demonstrating a more advanced capturing of the temporal structure in the gesture sequences. It seems that imitation learning tasks benefit more from transformer models than from conventional deep learning.