A Framework for Collaborative Digital-Physical Consistency Error Prediction in Multi-Robot Teams

Digital-physical fusion testing (DPFT) is a new testing method for complex equipment. It combines simulation testing of digital twin with real testing of physical equipment, thereby possessing characteristics of credibility and flexibility. At present, DPFT is a critical methodology for measuring the performance of multi-robot collaborative systems. However, the validity of DPFT hinges on maintaining high fidelity, which is often compromised by consistency errors between digital twin model and physical model. Traditional consistency error analysis methods are typically reactive or based on isolated models, failing to capture the emergent and coupled nature of errors in real-time collaborative scenarios. This paper introduces a framework for collaborative digital-physical consistency error prediction in multi-robot teams (C-DPCEP) designed specifically to enhance the reliability of DPFT. We reformulate consistency error prediction as a cooperative multi-agent reinforcement learning (MARL) task. The framework, named C-DPCEP, is based on a cooperative multi-agent reinforcement learning (MARL) paradigm and uniquely integrates a triple-attention mechanism into its policy network. This allows each robot agent to proactively identify error sources by focusing on its own state through self-attention, capture error propagation by evaluating peer states through inter-agent attention and account for external disturbances by assessing environmental entities through entity attention. The triple-attention mechanism can promote the algorithm to achieve higher prediction accuracy and faster convergence speed. By providing high-precision error predictions, the framework enables real-time compensation, significantly improving the fidelity and reliability of the testing process. This work provides a robust solution for ensuring the credibility of validation results for complex multi-robot collaborative algorithms.