Long-Term Adaptive Modeling for Autonomous Underwater Vehicles

Abstract

Autonomous Underwater Vehicles (AUV)s are robotic systems designed as an alternative to manned submerged vehicles, to carry out specific and well-defined underwater tasks. As part of their development process, modeling the hydrodynamicbehaviour of these vehicle is a crucial step for analysing their stability, implementing model-based control schemes, developing high fidelity simulators, and aidingtheir navigation systems. Models are generally developed by analysing the dynamics of the fluid surrounding the body. This involves estimating the motion states as function of their control inputs and other environmental aspects. However, infinite-dimensional analysis through the Navier-Stokes equations is computationally infeasible for any practical application, whereas finite-dimensional solutions usually sacrifice the model’s accuracy due to several assumptions and simplifications. Moreover, AUVs can generally succeed in performing short-term missions when the operating conditions are stable and their navigation sensors are in good condition. However, when the environmental situations change or their sensors fail, these robots can get easily confused or lost. This problem gets more severe when AUVs are required to operate for prolonged periods of time. Environmentally induced changes such as viscosity or density fluctuations, bio-fouling, body wear or damage and actuator malfunctions, are all factors that lead to a change in the robot’s expected behavior and affect directly their model’s integrity. This thesis tackles the aforementioned problems by developing automatic and efficient data-driven methods to construct AUV models from data streams accessible through the robot’s sensors. Motivated by advances in fields such as manipulator and humanoid robotics, we propose the use of model learning for underwater vehicles. Using data collected from real AUV experiments, we present first a comprehensive study that evaluates several learning methods and compare the results against classical physics-based approaches. Furthermore, we investigate methods to leverage from prior knowledge to improve the prediction accuracy of model learning as well as the number of samples required for training. Regarding the problem of time-varying dynamics due to environmental factors, we develop a framework to learn and adapt AUV models online. The proposed framework employs an incremental learning method to estimate the model sequentially from data streams available to the robot. In combination, strategies for including and forgetting data are developed to obtain better generalization over the whole statespace. The framework is tested in simulation and real experimental scenarios demonstrating its adaptation capabilities to changes in the robot’s dynamics. Moreover, we address the problem of identifying multiple contexts of an AUV model. We propose a gating network architecture that infers the different contexts directly from the data. Lastly, we address the issue of lifelong learning, where a robot can learn multiple tasks over a prolonged period of time. We present an architecture that can automatically learn and switch between different model contexts, while at the same time is capable of consolidating knowledge to learn new contexts and recall previously seen situations. Finally, the proposed approach is then validated in real experimental settings.