Riassunto analitico
Personalized object recognition in Embodied AI systems poses significant challenges, especially in unsupervised learning scenarios where no labeled data is provided. Real-world captured data, while providing photorealism and facilitating knowledge transfer from simulation to reality, is often limited by static scenes with fixed layouts. This static nature restricts the diversity of training data, hindering the development of adaptable models. To address this limitation, we propose a novel framework that integrates procedural object placement within photorealistic environments. Our approach dynamically generates diverse scene configurations and interactions, which enhances the variability of training data. This enables more robust and personalized learning of object representations, improving the adaptability of models in user-specific contexts.
Rather than training models from scratch, we tested existing foundational models to evaluate their suitability for personalized object recognition for Embodied AI. By assessing these models in procedurally generated environments, we identify those best suited for effective adaptation in personalized and dynamic tasks, establishing a baseline for future foundational models in this domain.
This work establishes a foundation for future research in Embodied AI, offering a scalable method for generating procedurally populated, photorealistic environments. Our approach enables the exploration of more realistic and dynamic settings, contributing a flexible dataset creation methodology and advancing the interaction capabilities of AI systems in real-world applications.
|
