Physics informed digital twin approach for real time predictive maintenance of collaborative robotic arms

Abstract This study proposes an architecture that employs an Agentic Edge Computing layer for offloading computationally intensive task processing from the cloud, thereby achieving sub-millisecond synchronization for multi-modal sensor data, including torque, vibration, and thermal patterns. To evaluate real-world performance, edge-native execution is conducted on an NVIDIA Jetson AGX Orin module for hardware-in-the-loop (HiL) validation. Hardware-in-the-loop (HiL) testing performed on the physical UR10e robot confirms that the Agentic P-DT maintains a 98.2% similarity with simulation responses, even under mechanical friction and stochastic workloads. A new Deep Reinforcement Learning (DRL) method is also developed at the edge to optimize resource allocation and predict Remaining Useful Life (RUL) under stochastic operational conditions. By embedding physical kinematic constraints within the neural network’s loss function, this approach shows a 28.4% relative improvement in fault detection sensitivity over traditional data-driven models. HiL validation results in dense industrial settings show that the P-DT approach preserves a 99.7% level of synchronization accuracy while lowering unplanned downtime by 34%. Additionally, the approach facilitates proactive self-correction capabilities to compensate for mechanical degradation by adjusting motion plans. These results offer a robust solution for the deployment of autonomous and self-healing robotic systems in latency-sensitive industrial settings.