Accurate prediction of interaction forces in robotic manipulators is critical for safe human-robot collaboration and precise manipulation. Multi-disk continuum robot arms present unique modelling challenges owing to high degrees of freedom, complex mechanical coupling, and non-linear force transmission. This study presents a investigation of eight regression algorithms Linear Regression, Ridge, Lasso, Decision Tree, Random Forest, Gradient Boosting, XGBoost, and Support Vector Regression applied to interaction force ( \(f_N\) ) prediction in 18-disk continuum arm configurations using squared Euclidean distance features. The dataset comprised \(n = 500\) unique arm configurations. All models were evaluated via 10-fold cross-validation with variance reporting, 95 % bootstrap confidence intervals, and Wilcoxon signed-rank statistical significance tests. On the single held-out test split, XGBoost achieved the lowest MSE while Random Forest tied for lowest MSE Cross-validated \(R^2\) values were near zero or negative across all models, confirming that spatial geometric features alone are insufficient for reliable force prediction. A feature-ablation study showed that squared-distance features outperform raw coordinates, and a bias–variance decomposition confirmed high irreducible bias as the primary performance bottleneck. These results are presented as a quantitative baseline identifying the limitations of static geometric features and charting a roadmap toward physics-informed and deep learning approaches.
