Abstract
<title>Abstract</title> <p>Tendon-driven continuum robotic arms are promising flexible manipulators for minimally invasive surgery, where reliable recognition of safe contact configurations is important for reducing excessive force, tissue irritation, and unintended injury. However, labeled interaction data from realistic surgical environments remain limited. This study presents an explainable and label-efficient machine learning framework for recognizing the interaction configuration type of a tendon-driven continuum robot. Using the publicly available End-Effector Interaction Force for Continuum Robotic Arms dataset, containing 750 samples from 150 contact points, we engineered tendon-, geometry-, contact-, and force-aware features. Supervised classification, semi-supervised learning, and unsupervised clustering were evaluated using group-aware data partitioning by contact point and ten-fold cross-validation. Explainability was assessed using tree-based feature importance, permutation importance, and Shapley additive explanations. CatBoost achieved the best supervised performance, with a mean test F1-score of 0.9888 and a cross-validated F1-score of 0.996 ± 0.006. Self-training with a Random Forest base learner achieved an F1-score of 0.955 using only 10% of the labeled training data and 0.986 using 50%. Unsupervised clustering did not naturally recover the configuration structure. Ablation analysis showed that normal force improved performance, while configuration identifier information mainly served as a protocol-leakage probe. These findings suggest that explainable, label-efficient learning may support safer contact-aware control strategies for continuum robots, although validation in tissue-like and surgical environments is still required.</p>