Despite the success of large-scale pre-training, Vision-Language-Action (VLA) models often exhibit limited generalization when deployed in novel settings. These models can fail when encountering shifts in visual domains, diverse environmental contexts, or variations in robotic embodiments. Such limitations hinder the practical application of foundation models in open-world robotics. This doctoral project aims to investigate the factors that restrict VLA generalization and develop strategies to improve their adaptability. The research will explore techniques such as cross-domain alignment, robust representation learning, and embodiment-agnostic architectures. The goal is to create robot learning frameworks that can reliably transfer knowledge across different tasks and physical platforms, ensuring consistent performance in unseen real-world scenarios.
