Modeling of AUV Dynamics with Limited Resources: Efficient Online Learning Using Uncertainty

TL;DR

This work addresses online learning of AUV dynamics under limited storage by using epistemic uncertainty to guide data rehearsal. It introduces three uncertainty-based data-selection strategies—Greedy, Threshold, and Threshold-Greedy—and evaluates them with three uncertainty estimators (Ensembles, MC Dropout, Flipout) on the Dagon AUV dataset. Results show that Threshold provides the most stable and effective performance, with the highest data-efficiency among the methods, while Ensembles generally offer the best uncertainty estimates. The findings suggest that uncertainty-guided online learning can improve robustness and extend the autonomy of AUVs by reducing storage and training demands, with practical deployment considerations discussed for future work.

Abstract

Machine learning proves effective in constructing dynamics models from data, especially for underwater vehicles. Continuous refinement of these models using incoming data streams, however, often requires storage of an overwhelming amount of redundant data. This work investigates the use of uncertainty in the selection of data points to rehearse in online learning when storage capacity is constrained. The models are learned using an ensemble of multilayer perceptrons as they perform well at predicting epistemic uncertainty. We present three novel approaches: the Threshold method, which excludes samples with uncertainty below a specified threshold, the Greedy method, designed to maximize uncertainty among the stored points, and Threshold-Greedy, which combines the previous two approaches. The methods are assessed on data collected by an underwater vehicle Dagon. Comparison with baselines reveals that the Threshold exhibits enhanced stability throughout the learning process and also yields a model with the least cumulative testing loss. We also conducted detailed analyses on the impact of model parameters and storage size on the performance of the models, as well as a comparison of three different uncertainty estimation methods.

Figures (9)

• Figure 1: Problem addressed in this work. Gathered data contains redundancy. To restrict the size of storage we determine whether points are useful for learning by quantifying uncertainty.
• Figure 2: Top view of the Dagon AUV, its degrees of freedom and related thrusters. The vehicle was fixed in the horizontal plane. It was controlled in surge, sway, and yaw ($u, v, r$) and steered by 3 thrusters $(n_1, n_2, n_3)$. Figure adapted from wehbe2017learning.
• Figure 3: Diagram of applied method. The online learning process can be augmented with uncertainty quantification. Based on uncertainty we can determine the usefulness of a point. Using it, we can choose to store and train on the incoming point, or whether to skip it.
• Figure 4: Input features of the Dagon dataset
• Figure 5: Targets of the Dagon dataset