The Dual-System Hierarchical Architecture: A Future Paradigm for Vision-Language-Action Models

With the breakthroughs of deep learning and large-scale pre-trained models in the fields of natural language processing (NLP) and computer vision (CV), Vision-Language-Action (VLA) models have opened up new opportunities for the development of embodied intelligence. This paper first reviews the limitations of traditional robotic manipulation methods—specifically, their heavy reliance on precise environmental modeling and the scarcity of training data—and explores the potential advantages of end-to-end large models for robotic control. We systematically categorize existing VLA approaches into four major classes: classic Transformer-based architectures, diffusion model-based methods, autoregressive prediction methods leveraging pre-trained large language models (LLM) or vision-language models (VLM), and hierarchical hybrid architectures that integrate LLM with specialized action experts. We highlight the architectural characteristics, strengths, and limitations of each category, trace the trend from monolithic end-to-end systems to layered dual-system designs, and propose key future research directions, including multimodal integration, expert routing mechanisms, and adaptive subtask decomposition for open-world scenarios.