Foundation Models for Construction Robotics: Transfer Learning and Vision-Language-Action Integration

This paper investigates the adaptation of pre-trained robotics foundation models, specifically RT-2 and PaLM-E, for construction automation applications. This research explores transfer learning methodologies across diverse construction tasks, develop vision-language-action (VLA) frameworks for instructable construction robots, implement few-shot learning approaches for rapid adaptation to novel operations, and design generalist robot policies capable of multi-task construction automation. The experimental results demonstrate that foundation models can achieve 78% task success rates with only 5-10 demonstrations, significantly outperforming traditional task-specific models. This research presents a comprehensive framework for deploying large-scale vision-language models in unstructured construction environments, achieving real-time performance with latencies under 200ms. The findings suggest that foundation models represent a paradigm shift in construction robotics, enabling unprecedented flexibility and adaptability in automated construction systems.