Vision-Language-Action Models: Foundations, Techniques and Applications

Vision-Language-Action (VLA) models mark a transformative breakthrough in embodied AI, seamlessly integrating visual perception, natural language understanding, and robotic control to create intelligent, adaptable agents. Leveraging pre-trained vision-language models (VLMs), VLAs harness Internet-scale semantic knowledge for robust generalization across novel tasks, including object manipulation, navigation, and human interaction. This paper offers a survey of VLA models, elucidating their core concepts, architectural designs, and wideranging applications. We examine key architectural elements such as multimodal encoders, fusion mechanisms, and action decoders. Applications encompass robotic manipulation, autonomous navigation, human assistance, and emerging areas including edge computing, industrial automation, healthcare, agriculture, and virtual reality. By synthesizing these advancements, we highlight VLAs' potential to transcend traditional boundaries in robotics, while addressing persistent challenges in tokenization, fusion robustness, and embodiment generalization. This work provides a reference for researchers and practitioners, emphasizing VLAs' promise in developing scalable, versatile systems for real-world embodied intelligence.