FOUNDATION MODELS AND THEIR ADAPTATION FOR VISUAL AND POLICY CONTROL IN GENERAL-PURPOSE ROBOTS

General-purpose robots are expected to operate autonomously across diverse, unstructured, and dynamically changing environments [1], [2].Traditional task-specific perception and control pipelines struggle to scale to such variability [2].Foundation models-large-scale models pretrained on massive multimodal datasets-have emerged as a promising paradigm to address these limitations by enabling robust visual understanding, semantic reasoning, and transferable policy learning [1], [3].This paper presents a comprehensive study of foundation models and their adaptation for visual perception and policy control in general-purpose robotic systems [1].We analyze the motivation for adopting foundation models in robotics, categorize visual and policy control approaches, and review recent advances in language, vision, and multimodal robotic foundation models [1], [3], [4].Furthermore, we discuss adaptation strategies including pretraining, alignment, fine-tuning, and reinforcement learning-based policy optimization [4], [5].Through an extensive literature review, we highlight how these models improve generalization, sample efficiency, and autonomy in real-world robotic deployments [2], [3].Finally, we outline open challenges, ethical considerations, and future research directions toward scalable and trustworthy robotic intelligence [3].