Authors: Alejandro Casado-Pérez, Marija Popović, Daniel Gutiérrez-Reina, Sergio L. Toral., Samuel Yanes
Organizations: Departamento de Ingeniería Electrónica, Escuela Técnica Superior de Ingeniería, Universidad de Sevilla, 41005 Sevilla, España · MAVLab, Faculty of Aerospace Engineering, TU Delft, The Netherlands
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.
Figures & tables
Fig. 1: Uncertainty-aware multi-agent IPP loop studied in this paper, shown at one step of a simulated mission. A fleet of four ASVs samples a physics-based oil spill simulation f(x) . The sparse measurements are passed to the information model, EDL, which returns in a single forward pass the reconstruction f^(x) (dashed: ground-truth spill contour) and its uncertainty σ(x) , subject to boundary conditions. The inset compares the error-uncertainty calibration of EDL with that of MC Dropout, Deep Ensembles and a GP. The planner uses the normalized σtotal to choose the next waypoints of each agent (dashed arrows) under the mission constraints, and therefore chooses the data the model receives at the next step.
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.
Fig. 3: Summary of the four planning strategies discussed: (a) Value-greedy planner, (b) Uncertainty-greedy planner, (c) R-H Orienteering, and (d) MCTS with uncertainty-adaptive branching.
Model
RMSE ↓
IoU ↑
NLL ↓
UCE ↓
Cal. ↓
σˉ
EDL
0.036
0.88
− 3.16
1.85
0.011
0.027
Deep Ensembles
0.084
0.61
− 2.53
8.13
0.084
0.080
MC Dropout
0.100
0.57
− 2.08
11.56
0.081
0.101
GP
0.118
0.38
8.41
11.33
0.129
0.074
TABLE I: Offline results on the 400 test spills (mean over spills), observed along random paths as in training. All metrics on the unobserved navigable cells with σ=σtotal . Cal.: mean deviation of the calibration curve from the diagonal (Sec. V-C ). UCE scaled by 103 . Best result in bold.
Planner
GP
MC Drop.
Deep Ens.
EDL
Value Greedy
7.2 (6.3)
9.3 (4.2)
24.0 (14.8)
0.7 (0.3)
Uncertainty-Greedy
4.0 (4.1)
11.4 (3.7)
23.3 (15.6)
0.8 (0.7)
ε -Greedy
5.6 (4.8)
12.4 (3.8)
15.9 (11.9)
0.9 (0.2)
RH Orienteering
1.9 (2.5)
10.7 (4.1)
31.6 (19.7)
0.8 (0.3)
MCTS
5.3 (16.2)
11.5 (4.0)
71.5 (48.4)
0.7 (0.2)
TABLE II: UCE per planner and model at the end of the mission, reported as mean (std) over 60 maps. All values are scaled by 103 ; lower is better. Best result per planner in bold.
Fig. 4: Calibration of the uncertainty estimates per model, one line per planner. For each prediction, σtotal and the absolute error ∣f(x)−f^(x)∣ on the unobserved cells are min-max normalized; cells are grouped into deciles of σ~ , and each point is the mean normalized error of a decile. Points on the diagonal indicate perfect calibration; above it the model is overconfident, below it underconfident.
Fig. 5: RMSE on the unobserved navigable cells at the end of each episode, per model and planner (log scale); 60 maps per model and planner.
Fig. 6: IoU between the reconstructed and the true spill regions ( τ=0.05 ) over the navigable cells at the end of each episode, per model and planner; 60 maps per model and planner.
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.
RMSE
IoU
B
EDL
MCD
DE
GP
EDL
MCD
DE
GP
50
0.031
0.102
0.083
0.123
0.88
0.52
0.58
0.35
100
0.024
0.086
0.097
0.110
0.93
0.58
0.69
0.41
150
0.023
0.081
0.122
0.085
0.92
0.60
0.77
0.55
300
0.020
0.067
0.135
0.054
0.93
0.70
0.88
0.78
TABLE III: MCTS with each model at the end of the mission, mean over maps (budgets 50-150: maps 60-89; budget 300: maps 0-59).
Fleet guided by
Planner
EDL
Deep Ens.
MC Drop.
GP
MCTS
0.0196
0.0200
0.0203
0.0240
RH Orienteering
0.0190
0.0192
0.0196
0.0229
Uncertainty-Greedy
0.0235
0.0292
0.0262
0.0260
Value-Greedy
0.0184
0.0251
0.0258
0.0220
ε -Greedy
0.0220
0.0233
0.0228
0.0237
TABLE IV: RMSE on the unobserved navigable cells at the end of the mission when every fleet’s observations are reconstructed by EDL; budget 300, mean over 60 maps.
Fleet
On spill
Excess error
p
own → other
own → other
EDL-guided MCTS
0.54 → 0.34
0.113 → 0.068
3⋅10−6
DE-guided MCTS
0.56 → 0.41
0.116 → 0.064
9⋅10−6
EDL-guided Value-Gr.
0.82 → 0.41
0.099 → 0.061
8⋅10−8
MCD-guided RH Orient.
0.65 → 0.44
0.062 → 0.045
0.008
GP-guided MCTS
0.24 → 0.24
0.042 → 0.045
0.57
TABLE V: Mask transplant (budget 300, 60 maps): each fleet’s observation mask applied to its own spill and to another map. Excess error: RMSE of the Deep Ensembles minus RMSE of EDL on the unobserved cells (mean over maps); p : Wilcoxon paired by map.
Model
Mean (ms)
Median (ms)
Train (min)
EDL
4.4
3.9
1.7
Deep Ensembles
18.5
16.2
8.0
MC Dropout
136.3
137.3
1.7
GP
4215
2851
–
TABLE VI: Prediction time per step on the same 120 observation sets (deep models on GPU, GP on CPU) and training time (50 epochs, 1600 spills, RTX 3090).
Informative path planning chooses robot routes and sensing actions that gather useful information while respecting travel, energy, time, and safety constraints.
Informative Path Planning for scalar field reconstruction uses predictive uncertainty to direct sensing vehicles toward maximally informative locations. Gaussian Processes provide this signal but their stationary isotropic kernels are misspecified for non-homogeneous phenomena such as oil spills, producing miscalibrated estimates that degrade planning. We investigate whether replacing the Gaussian Process with a well-calibrated Deep Ensemble improves path planning outcomes, and whether uncertainty quality interacts with the choice of planning algorithm. Five strategies (ε-Greedy, Value Greedy, Uncertainty Greedy, Monte Carlo Tree Search, and Receding Horizon Orienteering) share a common Deep Ensemble backbone trained on physics-based oil spill simulations. On held-out stochastic spill scenarios, the Deep Ensemble reduces normalised reconstruction error by 83% relative to the Gaussian Process baseline. Crucially, well-calibrated uncertainty amplifies the importance of the planning strategy: the performance gap between algorithms is negligible under miscalibrated models but becomes substantial under the ensemble, where multi-step lookahead planners outperform greedy selection by up to 32% in reconstruction error and achieve IoU above 0.85. Monte Carlo Tree Search is the recommended planner, matching Orienteering in reconstruction quality at an order-of-magnitude lower computational cost.
Samuel Yanes Luis, Alejandro Casado Pérez, Alejandro Mendoza Barrionuevo +3
Department of Electronic Technology, University of Sevilla, Sevilla, Spain · Department of Electronic Engineering, University of Sevilla, Sevilla, Spain
Accurate reconstruction of environmental scalar fields from sparse onboard observations is essential for autonomous vehicles engaged in aquatic monitoring. Beyond point estimates, principled uncertainty quantification is critical for active sensing strategies such as Informative Path Planning, where epistemic uncertainty drives data collection decisions. This paper compares Gaussian Processes, Monte Carlo Dropout, Deep Ensembles, and Evidential Deep Learning for simultaneous scalar field reconstruction and uncertainty decomposition under three perceptual models representative of real sensor modalities. Results show that Evidential Deep Learning achieves the best reconstruction accuracy and uncertainty calibration across all sensor configurations at the lowest inference cost, while Gaussian Processes are fundamentally limited by their stationary kernel assumption and become intractable as observation density grows. These findings support Evidential Deep Learning as the preferred method for uncertainty-aware field reconstruction in real-time autonomous vehicle deployments.
Samuel Yanes Luis, Alejandro Casado Pérez, Alejandro Mendoza Barrionuevo +3
Department of Electronic Technology, University of Sevilla, Av. de la Reina Mercedes S/N, 41012, Sevilla, Spain. · Department of Electronic Engineering, University of Sevilla, Av. de los Descubrimientos S/N, 41005, Sevilla, Spain.
Autonomous surface vehicles (ASVs) offer an efficient solution for environmental cleanup as well as search and rescue operations in open waters. Targets in these settings drift continuously, so efficient search must balance exploration of unobserved regions with tracking of known targets. However, most target tracking and pursuit scenarios consider simple guidance behaviours and short-term predictions for decision-making. In this letter, we address the problem of search and capture of multiple drifting targets, such as litter, in marine environments, using an adaptive informative planning approach. A key aspect of our strategy is a spatiotemporal informative planning method based on model predictive path integral (MPPI) control, a sampling-based model predictive control method. Our planner balances search and tracking objectives with a cost design that couples spatiotemporal target drift predictions and exploration objectives. The planner directly generates kinematic-level commands by optimising continuous trajectories over long horizons. In the interception stage, we use pure pursuit guidance for physical capture of the targets. Simulations show that our planner captures up to 18% more targets than the strongest baseline, proving its efficiency for environmental cleanup. Finally, we validate our approach in field trials with an ASV.
Sanjeev Ramkumar Sudha, Marija Popović, Erlend M. Coates
Norwegian University of Science and Technology (NTNU), Norway · TU Delft, The Netherlands