Mobile Low-Dwell Time Robotic Radiation Mapping Using Auto-Regressive Multi-Fidelity Gaussian Processes (ARGP) in Saturated Radiological Environments
Article
Cintas, B, Adams, J, Rios, C et al. (2026). Mobile Low-Dwell Time Robotic Radiation Mapping Using Auto-Regressive Multi-Fidelity Gaussian Processes (ARGP) in Saturated Radiological Environments
. Nuclear Technology, 10.1080/00295450.2026.2726706
Cintas, B, Adams, J, Rios, C et al. (2026). Mobile Low-Dwell Time Robotic Radiation Mapping Using Auto-Regressive Multi-Fidelity Gaussian Processes (ARGP) in Saturated Radiological Environments
. Nuclear Technology, 10.1080/00295450.2026.2726706
This work presents an evaluation of a multi-fidelity Gaussian Process (GP) framework for fusing sparse, high-accuracy “ground-truth” radiation measurements with the dense, lower-accuracy data stream of a mobile robotic platform into a single, continuous dose-rate map with quantified uncertainty. The framework is structured in two stages: (i) an offline kernel-selection stage that fits an additive log-space discrepancy model against a fixed-location site control survey, and (ii) an online deployment stage that re-uses the offline low-fidelity GP as a Kennedy and O’Hagan auto-regressive prior over a fresh GP fit on the live robot stream. Survey data collected at a university research reactor drives the evaluation, where cross-validation across four candidate covariance kernels (Radial Basis Function, Matérn 3/2, Matérn 5/2, and Rational Quadratic) is used to assess predictive accuracy and uncertainty calibration against the control. Three results follow from treating the validation itself as an object of study. A paired bootstrap over the 21 control points shows the four kernels to be all but indistinguishable under leave-one-out cross-validation, with 34 of 36 paired differences in root mean square error containing zero. A ranking emerges only under repeated spatially-blocked cross-validation, where both Matérn kernels beat the alternatives on every trial and the ordering is close to the reverse of the leave-one-out one, quantifying how much spatial leakage that protocol admits. The model’s systematic under-prediction of high-dose zones, between 19% and 58% depending on trial and blocking radius, is shown to be substantially a property of the constant-variance likelihood rather than of data sparsity alone: substituting the detector’s Poisson counting variance removes the bias on one trial and reduces it on the others, at a cost in overall error. The selected kernel is then deployed on the robot during near real-time mapping, where the model produces continuous heatmaps of the predicted dose rate and its paired uncertainty, supporting situational awareness and conservative dose-bound planning. The primary contribution is a quantitative characterization of how the combined framework behaves on a real reactor cell, including which of its assumptions are load-bearing and which are not, and how its uncertainty output can guide conservative dose planning and active sampling.
