Abstract: In human-robot collaborative disassembly, manufacturing systems predominantly rely on fixed perception-cognition paradigms governed by pre-established algorithms. This reliance poses significant challenges in accommodating the flexible requirements of operators, which are inherently informed by experience and the dynamic collaborative environment. As a result, robotic path planning often fails, and decision-making stalls. To address this, an embodied augmented reality disassembly system for human-robot-environment integration is proposed. The system is grounded in embodied intelligence theory and features a "perception-cognition-execution" mechanism. By combining this mechanism with augmented reality technology, it enhances environmental perception and cognitive reasoning. A collaborative disassembly strategy for embodied augmented reality is designed; a local image attention model with context enhanced mechanism to generate adaptive image captioning; a self-optimizing cognitive reasoning method is developed by large language model tuning and inference mechanism; a robotic manipulation method is developed through augmented reality-based human-robot-environment data interaction. Three similarity metrics are constructed to evaluate the performance of embodied perception and cognition. Quantitative and qualitative experiments demonstrate the system’s feasibility and effectiveness in enhancing human-robot collaborative disassembly efficiency and adaptability.

Key words: human-robot collaborative disassembly, embodied intelligence, augmented reality, image captioning, large language model