Cooperative Learning-Based Adaptive Control for Autonomous Underwater Vehicle Teams Under Complete Model Uncertainty

Document Type

Presentation

Date of Original Version

3-27-2026

Abstract

Autonomous underwater vehicles (AUVs) are increasingly used in missions such as inspection and seabed mapping. These missions involve highly nonlinear, uncertain dynamics and often exceed the practical capability of a single vehicle. These challenges motivate multi-AUV teams that improve efficiency and mission reliability by distributing the workload and enabling task/role reassignment when an agent degrades or aborts. They also motivate learning-enabled adaptive control frameworks designed for such teams to operate under complete model uncertainty, learn unknown dynamics online, and adapt in real time to changing environmental conditions. To the best of our knowledge, existing learning-based work on AUV control has largely focused on single-vehicle scenarios, often using deep reinforcement learning (DRL) policies. These approaches typically learn locally, are validated mainly through empirical results, and do not provide a systematic mechanism for cooperative learning and knowledge consensus across agents. Moreover, prior studies generally do not offer theoretical guarantees that the learned knowledge reaches consensus or that learned parameters converge, especially at the team level. To address these gaps, we propose a novel cooperative learning-based adaptive control framework in which the agents execute the mission cooperatively while performing cooperative online learning of unknown nonlinear dynamics. In this method, all nonlinear dynamics—including inertial, Coriolis/centripetal, and damping effects—are treated as unknown and are approximated using radial basis function (RBF) neural networks (NNs). To achieve team-level consensus, each agent exchanges local state information and compact learning variables (NN weight estimates) with its neighbors. A key innovation is that learning is explicitly cooperative: neighbor-to-neighbor information exchange drives the learned parameters toward agreement, so information collected by one agent benefits the entire team rather than remaining local. Moreover, the learned knowledge is reusable, enabling new missions and role/position reassignment without relearning from scratch. The proposed method is supported by rigorous theoretical analysis. We establish uniform ultimate boundedness (UUB) of all closed-loop signals, prove exponential convergence of the tracking errors to a small neighborhood of zero, and show convergence of the learned parameters to a neighborhood of their optimal values, providing a direct theoretical guarantee on learning quality. To corroborate the theoretical results, we perform numerical simulations in MATLAB R2024b. We consider a leader-following mission with a team of four AUVs required to maintain a prescribed geometric configuration relative to the leader and to each other, while allowing communication only between neighboring agents. The results show that the agents accurately track the leader trajectory while preserving the desired configuration throughout the mission, and the closed-loop system remains stable with exponentially fast error convergence. We further evaluate the learning performance by monitoring the NN weight estimates, which converge toward their optimal values. In addition, by comparing the learned model with the underlying unknown nonlinear term used in the simulation, we confirm that the proposed scheme accurately identifies the unknown dynamics.

This document is currently not available here.

Share

COinS