Unsupervised Learning of Graph Embeddings for Object-agnostic Affordance Categorization
Ramakrishnan Raman, Vikram Kumar, Dimple Saini, Dhaval Rabadiya, Smruti Patre, Ramakrishnan Meenakshi · 2024
In the dynamic field of artificial intelligence, the unsupervised learning of graph embeddings introduces a transformative approach to object-agnostic affordance categorization, enabling machines to understand and categorize object functionalities without relying on predefined categories. This paper outlines a novel methodology that integrates graph theory and deep learning to autonomously identify and categorize affordances by leveraging the intrinsic structural patterns of data. Through meticulous preprocessing of diverse datasets and construction of graphs based on feature similarity, our approach employs the Node2Vec algorithm to generate dense vector space embeddings that preserve essential topological structures. We further enhance affordance categorization by applying the K-means clustering algorithm, organizing objects into distinct clusters based on their affordances without the need for labeled data. Our extensive experimental results demonstrate marked improvements in precision (0.92), recall (0.89), and F1-score (0.90) over conventional models, highlighting the scalability and adaptability of our model. This research not only advances the understanding of affordance perception in a label-free framework but also showcases the potential of graph embeddings in revolutionizing autonomous systems and robotics, setting a new benchmark for future developments in unsupervised learning strategies.