Bayesian Optimization for real-time γ-localization measurements with robotic platforms

An adaptive approach driven by Bayesian Optimization is described for applications where remote radiation measurements made with robots are constrained by stringent upper thresholds on the mass and power payload of the necessary instrumentation, as well as by the time window within which measurements must be made, ultimately affecting their quality and maximum area coverage. The algorithm presented in this paper is applied to a gimbal assembly the comprises a collimated cerium bromide detector to perform γ-localization. A Gaussian Process models the angular distribution of radiation, and measurement locations are dynamically selected via a composite acquisition function, which combines the Expected improvement and Uncertainty Confidence Bounds functions. Convergence is driven by monitoring the rate of change of predictions and associated mean uncertainty. This approach enables accurate characterization of radiation fields while requiring up to 85% fewer measurements than conventional raster-type scanning. Its performance is evaluated in simulation, using previously obtained datasets used as measurement look-up tables, and validated in turn with hardware implementation, real-time scans.