Embodied Digital Twin driven human-centric collaborative robot behavior: Cognitive inference of action strategies

Collaborative robot interacting autonomously with human operators through flexible behavioral strategies poses a significant challenge. The lack of sufficient attention to the status of human operators during autonomous collaboration has led to safety hazards, necessitating improvements in current robot behavioral strategy research. Digital twin (DT) can establish interactive ecosystems consisting of assets and human nodes, providing continuous optimization for industry and becoming a crucial means for human-centric approaches. In this context, we propose Embodied Digital Twin (EDT) and its driven cognitive inference of collaborative robot behavior strategies, achieving safe and flexible autonomous collaboration through three stages: human-robot communication, human-robot understanding, and collaborative intelligence. Our approach has several key features: firstly, EDT shifts focus from traditional task-oriented approaches to human-centric engineering methods, offering a digital twin architecture with human-centricity, system empathy, and collaborative adaptability. Secondly, we establish an embodied environment for robot learning and training to address complex and dynamic human behaviors, enabling deep perception through spatiotemporal multisensory semantic fusion. Finally, utilizing neural-symbolic cognitive reasoning based on Large Language Model Multi-Agent System (LLM-MAS), we achieve robot autonomous collaboration aligned with human intent. We validate our approach using retired lithium battery disassembly tasks. Experimental results demonstrate that our proposed solution overcomes safety hazards caused by task rearrangements, physiological fatigue, psychological aversion, and human individual differences, showcasing improved safety, smoothness, and efficiency in collaboration performance. This confirms EDT as a universal guideline providing higher performance for human-centered solutions.