Abstract: Gaussian processes (GPs) are becoming a standard tool to build terrain representations thanks to their capacity to model map uncertainty. This effectively yields a reliability measure of the areas of the map, which can be directly utilized by Bayes filtering algorithms in robot localization problems. A key insight is that this uncertainty can incorporate the noise intrinsic to the terrain surveying process through the GPs ability to train on uncertain inputs (UIs). However, existing techniques to build GP maps with UIs in a tractable manner are restricted in the form and degree of the input noise. In this letter, we propose a flexible and efficient framework to build large-scale GP maps with UIs based on Stochastic Variational GPs and Monte Carlo sampling of the UIs distributions. We validate our mapping approach on a large bathymetric survey collected with an AUV and analyze its performance against the use of deterministic inputs (DI). Finally, we show how using UI SVGP maps yields more accurate particle filter localization results than DI SVGP on a real AUV mission over an entirely predicted area.
Introduction
Underwater terrain mapping is a critical yet challenging task for autonomous underwater vehicles (AUVs). Operating in GPS-denied environments, these vehicles rely on dead reckoning (DR), which inevitably accumulates error over time. This leads to uncertainty not only in sensor measurements (e.g., depth) but also in the spatial coordinates $(x, y)$ where those measurements are recorded. Traditional mapping techniques often treat these input coordinates as deterministic, failing to account for the inherent navigation noise. While Gaussian Processes (GPs) provide a probabilistic framework for surface modeling, standard "Uncertain Input" (UI) methods are often computationally prohibitive for large-scale surveys. This paper presents a scalable approach using Stochastic Variational Gaussian Processes (SVGPs) to build fully-probabilistic maps that incorporate input uncertainty.
Methodology
The proposed framework addresses the limitations of traditional GP mapping through two primary innovations:
- Uncertainty Propagation: The system uses the sigmapoint method to propagate AUV navigation uncertainty and sensor noise into the 3D coordinates of sonar beams. This treats each data point as a distribution (Uncertain Input) rather than a fixed point.
- SVGP with Monte Carlo Sampling: To handle millions of data points, the authors utilize Stochastic Variational Gaussian Processes. By employing Monte Carlo sampling from the UI distributions during training minibatches, the model naturally learns to incorporate input noise. This approach is computationally efficient and agnostic to the specific type of noise distribution.
These probabilistic maps are then integrated into a Particle Filter (PF) localization algorithm, allowing the AUV to estimate its position by leveraging the map's uncertainty estimates.
Experimental Setup
The framework was validated using a real-world bathymetric dataset collected by a Hugin 3000 AUV off the Swedish west coast. The survey covered approximately 1.4 $km^2$ and generated over 4 million multibeam echosounder (MBES) beams over a 12.7-hour mission.
Results and Discussion
A key finding of the research is that accounting for input uncertainty significantly improves the model's ability to predict unsampled areas. While reconstruction of observed areas remains similar between Deterministic Input (DI) and UI models, the UI-SVGP produces a smoother, more generalized representation that better reflects the true terrain when data is noisy.
Quantitative results demonstrate the superiority of the UI approach:
- Prediction Accuracy: In predictive tasks, the UI-SVGP reduced the Root Mean Square Error (RMSE) to 0.566m, compared to 0.766m for the DI-SVGP.
- Localization Performance: When navigating over entirely predicted seabed areas, the UI-SVGP map enabled a lower position error (10.71m) than the DI-SVGP map (11.88m) and raw dead reckoning (12.32m).
- Computational Efficiency: The UI approach required only a marginal increase (4-6%) in training iterations to reach convergence, making it a viable solution for large-scale mapping.
Conclusion
By combining SVGPs with Monte Carlo sampling of uncertain inputs, this research provides a scalable and flexible solution for probabilistic underwater mapping. The resulting maps offer more reliable uncertainty estimates, directly translating to improved AUV localization accuracy. Future work will explore automated hyperparameter tuning for inducing points and the development of real-time, online mapping capabilities.
Summary updated