Probabilistic model predicts continuum robot configurations across designs
A Probabilistic, Rotation-Preserving Kernel Operator Surrogate for Continuum-Robot Design Spaces
RoboticsMachine Learning
Summary
Continuum robots, which bend and twist like tentacles, are hard to predict because their shapes depend on many design choices and how they are controlled. The authors created a computer model that learns how different robot designs relate to their final shapes, including uncertainty estimates about the results. Their model uses math tools that respect the robot’s rotation properties and improves accuracy with less training data compared to previous methods. This helps engineers quickly understand how changing robot designs will affect robot shapes without running expensive simulations every time.
What this means in practice
- •For robotics engineers: Predict continuum robot shapes accurately across different designs and actuation with uncertainty estimates for design optimization.
- •For mechanical design teams: Reduce the number of physical prototypes by using the model to explore design variations and their impact on robot configurations.
Authors
Branden Frieden, Alan Kuntz, Varun Shankar
Abstract
We develop an uncertainty-aware, operator-learning-based surrogate for tendon-driven continuum robots (TDCRs) that generalizes across robot design and actuation while predicting the complete equilibrium configuration. Our central idea is a nonstationary Gaussian-process (GP) kernel whose metric is learned by a shallow neural network from mechanics-aware design coordinates. The network learns how similarity changes across design space, while the resulting kernel defines the design-to-pose operator, its posterior uncertainty, and a criterion for sequential simulator acquisition. We represent backbone position and rotational strain with spatial kernels and reconstruct orientation directly on SO(3), preserving valid material frames and tendon geometry. We use ASkotch with Nystrom preconditioning for scalable kernel operator solves and a separate Nystrom approximation for posterior covariance. Experiments show that kernel-coverage acquisition reduces mean trajectory error by approximately 15% relative to IID sampling, while posterior variance increases for increasingly unsupported designs. Compared with prior neural-operator TDCR models, the method is 30-50% more accurate using approximately 5x fewer training designs while additionally providing posterior uncertainty estimates.