Researchers led by Alejandro Casado-Pérez have studied how fleets of autonomous surface vehicles can map an oil spill from sparse, noisy measurements, according to a paper on arXiv. The vessels coordinate online through a shared map of uncertainty and decide where to sample next.
The team proposes evidential deep learning (EDL) as the model that estimates the spill and its uncertainty. It compared EDL, Monte Carlo dropout and deep ensembles with a Gaussian process, each with five planners, on 60 physics-based oil spill simulations. The choice of model explained 80% of the variation in final mapping error, against 1% for the planner among the deep models.
EDL had the lowest error under every planner, was the best calibrated and ran at 3.9 ms per step. With 50 sampled cells per vehicle it mapped the spill better than the other deep models did with 150. Deep ensembles grew more confident as their error grew, while EDL did not.



