Machine Learning Falls Short at Predicting Robot Arm Contact Forces

Machine learning models trained on the geometry of a slender, disk-based continuum robot arm cannot reliably predict the forces the arm will exert when it touches a person or an object, according to an open-access study published on 24 June 2026 in Discover Robotics. The work, led by Jigneshkumar P. Desai, evaluated eight regression algorithms on interaction force prediction for an 18-disk continuum arm and arrived at an unusually candid conclusion: the spatial geometric features available in the dataset are simply not enough, and no amount of algorithmic tuning within the tested family fixes that.

Multi-disk continuum robot arms are built from rigid or semi-rigid circular spacer disks arranged along a flexible backbone, typically elastic rods or superelastic nickel-titanium tubes. Unlike conventional serial-link robots with discrete joints, these manipulators deform continuously along their length; each disk constrains the local bending plane and curvature, and together the disks define the arm’s three-dimensional shape. The platform studied here has a fixed base, 18 intermediate disks, an end-effector platform, and actuation tendons. Such slender, dexterous morphologies are attractive for surgical assistance, industrial assembly, rehabilitation, and collaborative workspaces, where robots must interact physically and safely with humans.

Modelling the forces these arms transmit is genuinely hard, and the paper lays out why. With 18 disks each described by three spatial coordinates, the configuration space is 54-dimensional. Tendon routing couples non-adjacent disks, so a local geometric change propagates non-locally through the structure. The elastic backbone exhibits non-linear stress-strain behaviour, and hysteresis in tendon-sheath systems, disk friction, and viscoelastic creep further complicate the mapping from shape to force. On top of that, interaction forces depend on the geometry and properties of whatever object is contacted, and force-torque sensors suffer from noise, drift, and cross-axis coupling. Traditional physics-based models such as Cosserat rod theory or the piecewise constant curvature approximation demand precise material identification and are expensive to evaluate at run time, which is why data-driven machine learning, capable of microsecond inference once trained, is an appealing complement.

The study assembled a dataset of 500 unique arm configurations after aggregating by configuration type, split 80 percent for training and 20 percent for testing with a fixed random seed. Each sample recorded the central reference point, the positions of all 18 disks, a configuration-type grouping key, and the measured interaction force, which had a mean of 0.2367 newtons, a standard deviation of 0.1233 newtons, and a range from minus 0.1758 to 0.5614 newtons. The feature engineering was deliberately simple: for each disk, the squared Euclidean distance from the disk to the central reference point was computed, yielding 18 scalar features that encode each disk’s radial deviation from the arm’s central axis. Squared distances avoid a square-root computation, are monotone in distance ordering, and are invariant to rigid-body rotations and translations.

Eight regression algorithms faced off: Linear Regression, Ridge, Lasso, Decision Tree, Random Forest, scikit-learn’s Gradient Boosting, XGBoost, and Support Vector Regression. Random Forest and XGBoost were tuned through 5-fold inner cross-validation grid search, while the other models ran on default parameters to keep the initial comparison unbiased. Evaluation was unusually thorough for a baseline study: mean squared error and the coefficient of determination were reported on a single held-out split with 95 percent bootstrap confidence intervals built from 1,000 resamples, 10-fold cross-validation with variance across folds, and Wilcoxon signed-rank significance tests comparing each model’s per-fold performance against the best performer.

The headline numbers are sobering. On the single test split, XGBoost achieved the lowest mean squared error, 0.0153, and the only positive R-squared value, 0.0199, while Random Forest recorded a mean squared error of 0.0156 with an R-squared of essentially zero. Six of the eight models returned negative R-squared, meaning they performed worse than a predictor that simply outputs the dataset’s mean force. Under 10-fold cross-validation the picture did not improve: Random Forest posted the best cross-validated mean squared error, 0.0150 plus or minus 0.0023, and the least-negative cross-validated R-squared, minus 0.011 plus or minus 0.075, but every model sat at or below the constant-mean predictor across the folds. The authors state plainly that cross-validated R-squared values near zero or negative across all models confirm that spatial geometric features alone are insufficient for reliable force prediction.

The statistical tests added nuance rather than rescue. Wilcoxon signed-rank comparisons against Random Forest found significant differences for the Decision Tree, with a p-value of 0.002, XGBoost at 0.020, and Support Vector Regression at 0.037, while the linear models and Gradient Boosting did not differ significantly, indicating they share similar, very limited explanatory power under this feature set. Large fold-to-fold standard deviations, notably an R-squared standard deviation of 0.446 for the Decision Tree, underscored why cross-validation matters before declaring any winner. A validation heatmap summarising every model and metric showed XGBoost leading on test-set scores while Random Forest led on cross-validated consistency, a divergence the authors attribute to different bias-variance trade-offs in this near-zero regime.

Two diagnostic analyses pinpointed the root cause. A feature-ablation study on Random Forest compared the 18 squared-distance features against all 54 raw coordinates under 10-fold cross-validation: the squared distances achieved an 8 percent lower mean squared error and a better R-squared, minus 0.011 versus minus 0.068, an advantage credited to rotational invariance, yet both feature sets remained near zero in explanatory power, showing the feature type is a secondary issue compared with what is missing. A bias-variance decomposition over 30 repeated random 80/20 splits found that squared bias dominates variance for the regularised linear models and the ensemble methods, the signature of systematic underfitting, while the unpruned Decision Tree showed the highest variance. Feature-importance analysis from the tuned Random Forest showed the disks nearest the tip carrying the greatest weight, consistent with the mechanical intuition that tip displacement most directly governs end-effector interaction forces. Residual distributions for all models were approximately centred at zero, indicating no systematic directional bias in the errors.

The authors are explicit about what the negative scores mean and what the study does not do. Negative R-squared values, they write, explicitly confirm model failure on this feature set, but the failure is a feature-space limitation rather than an algorithmic one: geometric distances omit friction coefficients, hysteresis states, tendon tensions, velocities, accelerations, and material stiffness, all of which are critical to determining interaction forces. The evaluation was offline only, with no hardware deployment, no real-time inference benchmarking, and no noise or robustness experiments. No deep learning comparators, such as convolutional, recurrent, transformer, or physics-informed neural network baselines, were included. The dataset of 500 configurations is also modest by modern machine learning standards.

What the paper offers instead is a rigorous, honest baseline and a roadmap. Future work, the authors argue, should incorporate dynamic and actuation features such as tendon tensions, joint velocities, and accelerations; combine Cosserat rod constraints with data in physics-informed machine learning; benchmark deep learning architectures; validate with hardware-in-the-loop force sensors; and test robustness under domain shift and sensor noise. Until then, the lesson for the human-robot collaboration community is concrete: for disk-based continuum arms, richer sensing of the arm’s internal state, not a smarter regressor, is the path to force predictions trustworthy enough for safe physical contact.

Subject of Research: Machine learning prediction of interaction forces in multi-disk continuum robot arms

Article Title: Machine learning based interaction force prediction in multi-disk continuum robot arms for human-robot collaboration

Article References: Desai, J. P. (2026). Machine learning based interaction force prediction in multi-disk continuum robot arms for human-robot collaboration. Discover Robotics, 2(1), Article 19. https://doi.org/10.1007/s44430-026-00033-5

Image Credits: AI Generated

DOI: 10.1007/s44430-026-00033-5

Keywords: continuum robots, human-robot collaboration, interaction force prediction, machine learning, XGBoost, random forest, cross-validation, bias-variance decomposition, feature engineering, squared Euclidean distance, regression algorithms, robotic manipulators