Compensating control delays in the CFETR Multi-Purpose Overload Robot for divertor maintenance by using transformer-enhanced Temporal Convolutional Network

Communication delays in the CFETR (China Fusion Engineering Test Reactor) Multi-Purpose Overload Robot (CMOR), designed for divertor maintenance, significantly degrade control precision and operational reliability. This study proposed a Transformer-enhanced Temporal Convolutional Network (TransformerTCN) to address these delays. By integrating the local temporal feature extraction of Temporal Convolutional Networks (TCNs) with the long-range dependency modeling of Transformer’s self-attention mechanism, TransformerTCN accurately predicted future robot states and dynamically adjusted control commands. Validation in ROS-based simulations demonstrated that TransformerTCN outperformed traditional models, achieving a Mean Squared Error (MSE) of 0.002 and an $R^2$ score of 0.994, showcasing its superior accuracy and robustness. These results highlight TransformerTCN as a powerful solution for delay compensation in robotic systems operating in challenging nuclear fusion environments.