Uncertainty Estimation for Deep Reconstruction in Actuatic Disaster Scenarios with Autonomous Vehicles

A new research paper from a team of six scientists, including Samuel Yanes Luis and Sergio Toral Marín, provides a critical comparison of AI methods for a core robotics task: reconstructing an environment from sparse sensor data while accurately quantifying uncertainty. This capability is essential for autonomous vehicles (AVs) operating in complex, dynamic scenarios like aquatic disaster zones, where they must map pollutants or hazards. The study rigorously evaluated four prominent techniques—Gaussian Processes (GP), Monte Carlo Dropout, Deep Ensembles, and Evidential Deep Learning (EDL)—against three perceptual models mimicking real-world sensors.

The findings are decisive for practical deployment. Evidential Deep Learning emerged as the preferred method, delivering the highest reconstruction accuracy and the best-calibrated uncertainty estimates across all tested sensor configurations. Crucially, it did so at the lowest computational inference cost, a key factor for real-time operation on vehicle hardware. In contrast, while theoretically sound, Gaussian Processes were hampered by their stationary kernel assumption and became computationally intractable as data density increased. This research provides a clear, evidence-based guideline for engineers building the next generation of AVs for environmental monitoring and disaster response, prioritizing both performance and efficiency.