A Probabilistic Learning Framework for Forward and Inverse Kinematics of Concentric Tube Robots
Year: 2026
Type: article
Abstract: Accurate kinematic models are essential for motion planning and closed-loop control of concentric tube robots (CTRs); however, experimental CTR behavior exhibits both history-dependent forward and intrinsically multi-valued inverse kinematics, making purely memoryless deterministic models inadequate for capturing the observed input-output relationships. This work addresses both challenges using separate mixture density networks (MDNs), which model each mapping as a conditional probability distribution rather than a deterministic function. For forward kinematics, the MDN input is augmented with a memory buffer of recent joint increments to capture short-term path dependence, with model selection based on validation negative log-likelihood (NLL). The history-augmented model achieves sub-millimetric mean tip-position error of 0.63 mm, 0.30% with orientation error ofmmusing the maximum-a-posteriori component, and FK-guided candidate selection further reduces this error to 1.85mmby selecting among multiple posterior inverse candidates. Overall, the results demonstrate that probabilistic MDN-based models provide an effective framework for representing both forward path dependence and inverse-solution multiplicity in CTR kinematics.
Institution: University of Naples Federico II
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7223263340
