Physics-informed AI lets robots find radiation sources without risky path planning
New physics-guided ML model locates radiation sources from any robot path—no approach needed.
A team of researchers (Son, Tan, and Zhang) has introduced a physics-guided machine learning framework for robotic radiation source localization (RSL) that breaks away from traditional path-planning constraints. Existing methods require robots to approach the source for precise estimation, increasing radiation damage risk and limiting mission flexibility. The new approach uses physics-informed neural networks with specially designed tensors that model gamma-ray flux attenuation caused by unknown obstacles. Multiple models run in parallel to boost robustness and precision, enabling accurate source localization regardless of the robot's measurement path.
Evaluation included high-fidelity Monte Carlo particle transport simulations across diverse environments—varying spatial scales, source types, obstacle materials, and robot trajectories. The method was also validated in physical experiments using configurations not present in training data. A continuous learning technique further enhances performance during real robot deployment. The authors describe this work as advancing robot radiation perception from pointwise flux detection to spatial intelligence, allowing robots to operate safely in radioactive zones without rigid path planning.
- Physics-informed ML handles attenuated gamma-ray signals from unknown obstacles without requiring obstacle geometry knowledge.
- Allows arbitrary robot measurement paths, eliminating the need to approach the radiation source and reducing damage risk.
- Validated in both high-fidelity Monte Carlo simulations and real-world physical experiments with unseen configurations.
Why It Matters
Safer, more flexible robotic radiation mapping for nuclear facilities and disaster response without special path planning.