Physics-Guided Robotic Radiation Source Localization along Arbitrary Measurement Paths in Unstructured Environments

Hojoon Son, Kai Tan, Fan Zhang

Using robots to estimate the location of the radiation source is an effective way to improve efficiency and safety. Existing methods focus on planning the robot's path to achieve precise estimation, typically approaching the source. However, approaching the source increases the risk of radiation damage to a robot. In addition, a path-planning algorithm designed solely for radiation source localization (RSL) limits the flexibility of missions that deploy robots into radioactive environments. This study presents an automation framework for robotic RSL that leverages a physics-informed machine learning (PIML) model to precisely estimate the source location, regardless of measurement paths, in unknown environments. Physics-inspired model tensors have been designed for PIML to handle attenuated gamma-ray flux signals from unknown obstacles, and multiple models are computed in parallel to improve the robustness and precision of the RSL. The proposed method is evaluated in high-fidelity simulation environments using Monte Carlo particle transport across diverse randomized domains, including spatial scales, radiation source types, obstacle materials and geometries, and robot trajectories. The method is also validated through physical experiments on configurations that are not included in the simulation-based evaluation. The continuous learning technique is applied in real-robot deployment to enhance the practical applicability of the online robotic RSL system. The proposed method advances robot radiation perception from pointwise flux detection to spatial intelligence.