Configuration-constrained data-driven inverse kinematics for continuum manipulators: Bayesian-regularized learning and uncertainty-aware validation
Year: 2026
Type: article
Abstract: Multi-section continuum manipulators may admit several internal configurations for the same task-space target, which can generate inconsistent labels for data-driven inverse kinematics. This paper proposes a configuration-constrained data-generation framework for a two-section continuum manipulator. Each section is discretized into routed spacer-disk elements, and actuator-length labels are computed from the resulting geometric path. The one-to-one mapping is used in a restricted dataset sense: under constant curvature, fixed routing, bounded actuation, no slip, no load, and the selected-configuration rule, each retained Cartesian sample is assigned one stored actuator-length vector. No global physical bijection is claimed under friction, hysteresis, gravity, payload, or non-constant-curvature deformation. Three regressors are then compared: Bayesian Regularized Deep Neural Network (BR-DNN), Random Forest, and XGBoost. Errors are interpreted separately as actuator-space model-reference errors, task-space reconstruction errors, experimental uncertainty, and inference latency. On the held-out trajectory generated from the same analytical family, BR-DNN gives the lowest actuator-space RMSE, approximately 0.14 mm, with smoother predictions than the tree-based models. The Robotino®XT/CBHA experiment is used as an external consistency check based on equivalent actuator-length/tube-elongation signals. Because the combined measurement uncertainty is about ± 0.27 mm, sub-millimeter numerical values are interpreted cautiously. The single-query inference time supports high-rate command generation feasibility, while complete closed-loop bandwidth remains a future hardware-level task.
Source: Control Engineering Practice
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,010
OpenAlex ID: W7204586935