Better Calibrated, Better Planned: Uncertainty-Aware Multi-ASV Informative Path Planning for Oil Spill Monitoring

Alejandro Casado-Pérez, Marija Popović, Daniel Gutiérrez-Reina, Sergio L. Toral., Samuel YanesarXiv 2026 · arXiv

TL;DR

In the high-stakes task of monitoring oil spills with robot fleets, the "brain" matters more than the "legs." This paper reveals that the accuracy of an informative path planning (IPP) system depends far more on how well a model quantifies its own uncertainty (calibration) than on the complexity of the path planning algorithm itself. By introducing Evidential Deep Learning (EDL) as a fast, well-calibrated information model, the authors achieve significantly better spill reconstruction with a fraction of the computational cost of traditional methods.

What Problem Does This Paper Address?

When an oil spill occurs, a fleet of Autonomous Surface Vehicles (ASVs) must map the pollutant field quickly using sparse, noisy sensor readings. This is a classic Informative Path Planning (IPP) problem: the robots must decide where to move next to maximize the information they gather.

Traditionally, researchers have relied on Gaussian Processes (GPs) to model the field and its uncertainty. However, GPs are computationally expensive—their cost explodes as more data is collected—and their uncertainty estimates are often "blind" to the actual values measured, relying mostly on how far a robot is from previous sample points. Deep learning models offer a faster alternative, but they often suffer from poor calibration: they might be very confident in a prediction that is actually wrong. In a closed-loop system, a miscalibrated model misleads the planner, causing the robots to ignore critical areas or waste time in uninformative ones.

Methods

The authors propose using Evidential Deep Learning (EDL) as the core information model. Unlike standard neural networks that output a single value, EDL predicts the parameters of a probability distribution (specifically, a Normal-Inverse-Gamma distribution). This allows the model to estimate both aleatoric uncertainty (noise in the data) and epistemic uncertainty (the model’s lack of knowledge) in a single forward pass.

The sensing-learning-planning loop using EDL.

The study benchmarks EDL against three other models:

  1. Gaussian Processes (GP): The traditional baseline.
  2. Monte Carlo Dropout (MC Dropout): A deep learning method that estimates uncertainty by running multiple passes with random neurons turned off.
  3. Deep Ensembles: Training multiple independent models and measuring their disagreement.

These models were paired with five different path planners, ranging from simple "Greedy" strategies to complex look-ahead algorithms like Monte Carlo Tree Search (MCTS) and Receding Horizon (RH) Orienteering.

Fig. 2. Uncertainty estimation methods compared in this work: MC Dropout, Deep Ensembles and EDL. All three share the same U-Net backbone (top); they differ only in how the backbone is used and in the output heads attached to it.

Datasets

The researchers used a physics-based simulator to generate 2,000 unique oil spill scenarios on a 100x100 grid.

  • Training: 1,600 spills were used to train the deep learning models.
  • Testing: 400 held-out spills were used for offline evaluation.
  • Closed-loop Evaluation: 60 maps were used to test how the models performed when actually guiding a fleet of four ASVs in real-time.

Key Insights and Ideas

  • The Model Dominates the Planner: In a striking finding, the choice of the reconstruction model explained 80% of the variance in final mapping error, while the choice of the path planner explained only 1%. This suggests that researchers should focus more on model reliability than on hyper-optimizing path trajectories.
  • The "Closed-Loop" Trap: Planners naturally steer robots toward areas of high uncertainty. If a model is poorly calibrated (e.g., it becomes overconfident in its mistakes), the planner will "blindly" avoid the very areas it needs to sample.
  • Deep Ensembles Can Fail Upward: While Deep Ensembles are often considered the gold standard for deep uncertainty, the authors found they actually became more confident as their error grew during a mission. This is because the data collected by the planner shifted away from the training distribution, a phenomenon known as distribution shift.
  • EDL is Robust: EDL remained well-calibrated even as the robots concentrated their samples on the spill, making it uniquely suited for the feedback loop of autonomous exploration.
Calibration curves showing EDL's superior performance across different planners.

Key Results

  • Superior Accuracy: EDL achieved the lowest reconstruction error (RMSE) and the highest Intersection over Union (IoU) across every single planner tested.
  • Efficiency: EDL reached a high level of accuracy within the first 20% of the mission budget, outperforming what other models achieved at the very end of their missions.
  • Speed: EDL is roughly 700 times faster than a GP at the median mission step, taking only 3.9 ms per prediction. This makes it highly suitable for real-time deployment on low-power hardware.
  • Sample Efficiency: With only 50 samples per vehicle, EDL reconstructed the spill better than MC Dropout or Deep Ensembles did with 150 samples.
Fig. 7. RMSE on the unobserved navigable cells along the mission, MCTS with budget 300, mean over 60 maps; each model reconstructs the observations of its own fleet. Dotted line: 20% of the budget.

Limitations and Future Work

While the results are promising, the authors note several areas for further research:

  • Static Fields: The current study assumes the oil spill is static. In reality, spills drift and spread over time, requiring models that can handle temporal dynamics.
  • Communication & Localization: The simulation assumes perfect communication between robots and exact GPS positioning, which are rarely guaranteed in maritime environments.
  • Multi-Objective Planning: Future work could explore planners that balance the need for accuracy with other constraints, like battery life or the need to "geofence" (delineate the boundaries) of the spill specifically.

Summary updated

Abstract

Monitoring oil spills with fleets of Autonomous Surface Vehicles (ASVs) requires mapping a heterogeneous pollutant field from sparse, noisy measurements under tight time and energy budgets. Informative Path Planning (IPP) addresses this problem by coupling a reconstruction model, which infers the field and its uncertainty, with a planner that decides where to sample next. Prior work has focused on planning strategies and has largely relied on Gaussian Processes (GPs), leaving the calibration of the model's uncertainty underexplored. We study how uncertainty calibration conditions multi-vehicle IPP with a fleet of ASVs that coordinates online through a shared uncertainty map. We introduce Evidential Deep Learning (EDL) as an uncertainty-calibrated information model. We benchmark EDL, Monte Carlo Dropout and Deep Ensembles against a GP in closed loop with five planners on physics-based oil spill simulations. Across 60 maps, the choice of model dominates the outcome, explaining 80\% of the variance of the final reconstruction error, against 1\% for the planner among the deep models. EDL achieves the lowest error and the highest IoU under every planner, is the best calibrated model and runs at 3.9 ms per step. With 50 cells per vehicle it reconstructs the spill better than MC Dropout and Deep Ensembles with 150. Deep Ensembles become more confident as their error grows, whereas EDL is unaffected. Robustness of the information model to the data its own planner collects is therefore a requirement of uncertainty-aware IPP.
BibTeX
@misc{casadoperez2026better,
  author = {Alejandro Casado-Pérez and Marija Popović and Daniel Gutiérrez-Reina and Sergio L. Toral. and Samuel Yanes},
  title = {{Better Calibrated, Better Planned: Uncertainty-Aware Multi-ASV Informative Path Planning for Oil Spill Monitoring}},
  year = {2026},
  eprint = {2611.21783},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21783},
}
Figure