Geometric-Aware Attention Networks with Adaptive Constraint Adjustment for Static Modeling of the Tendon-driven Continuum Robot

Tendon-driven continuum robots (TDCRs) with hyper-redundant DOFs are widely developed for sophisticated tasks such as minimally invasive surgery. However, the static modeling of the TDCRs is challenging due to the nonlinear characteristics of the TDCRs’ structures and material properties. To tackle the above-mentioned challenges, we develop a novel end-to-end deep learning approach called geometric-aware attention network (GAANet) by combining neural network architecture and domain-specific geometric constraints. To acquire the datasets and evaluate the applicability of our developed network, we built up a two-segment TDCR with variable cross-sections, which is equipped with load cells and a motion capture system to measure the tension of tendons and the tip position of each segment simultaneously. To prove the efficacy of the developed GAANet, a multilayer perceptron (MLP) trained with the same dataset and model-based approach called pseudo rigid body model (PRBM) are both implemented for comparison. The results show that the developed GAANet outperforms the MLP and PRBM by 70.33% and 73.36% respectively in terms of the tip position prediction accuracy of each segment, demonstrating its potential for precise and real-time static modeling for our developed TDCRs with variable cross-sections.