VLAbot: A human Vision–Language–Action models interaction framework for robotic assembly

In this work, we propose an intelligent human–robot collaboration system designed to assist embodied intelligence in learning complex, long-horizon manufacturing assembly tasks. The system integrates multiple expert agents and augmented reality (AR) interaction interfaces, enabling robots to request planning and execution guidance from humans and efficiently complete intricate tasks. Specifically, the expert agents equipped with a high-level planner and a Vision-Language-Action (VLA) model actively interact with users through text, vision, and action modalities to acquire critical information, learn task-specific skills, and develop sub-task planning strategies. A distributed data and model architecture ensures real-time interactions between different models and facilitates seamless human–robot collaboration. We evaluate the system on two challenging long-horizon manufacturing assembly tasks (gear assembly and peg insertion) to demonstrate the effectiveness of the proposed approach. The system successfully learns both assembly tasks within five trials and enables the embodied intelligence to complete them with progressively reduced execution time.