Gaussian processes applied to system identification, navigation and control of underwater vehicles

Autonomous underwater vehicles (AUVs) are increasingly being used in commercial, military and scientifc organisations for various applications ranging from hull inspections to underwater explorations. Despite the steady growth in applications, communication with the AUVs has been the major challenge, mainly due to the cost and limited availability of underwater communication system. Therefore, AUVs are required to have a higher degree of autonomy compared to the land based and aerial autonomous vehicles. In addition to the lack of communication, AUVs cannot measure their position or orientation with high accuracy since many states are calculated or corrected based on internal observers which employ other measurements and mathematical models. Popular methods of motion control and inertial navigation for AUVs require costly experiments as the mathematical representation of the vehicle is used for the predction and correction of states, controller deduction and the tuning procedure. New non-parametric algorithms for inertial navigation and motion control are required as the standard techniques employ a very complex mathematical model of the vehicle that usually is just an approximation as it is calculated based on ideal conditions. New non-parametric models can lead to the improvement of position estimation and robust control of a vehicle as these models are estimated over real data from the platform and they can include models of perturbances. Machine learning has become a principal tool in robotic applications as drones, autonomous vehicles and earth bioinspired robots. A specific method of machine learning that can be employed for the construction of non-parametric models is Gaussian Processes (GPs). GPs are based on the theory that natural phenomena can be fitted to a multivariable Gaussian distribution. This research started with an investigation to check whether Multi-output Gaussian Processes are capable of producing a non-parametric GPs model for ships and underwater vehicles with the advantage of being able to represent the relation between the outputs. In a similar way, it was researched whether GPs can be employed for post-processing of pressure sensors data to estimate the vehicle speed without other types of sensors. Another application explored was the application of GP-UKF (Gaussian processes unscented Kalman Filter) for navigation. GP-UKF was successfully simulated to navigate an underwater vehicle without the need to recognize a mathematical model as the model used is a GPs model identified in previous experiments. Finally, GPs were employed to learn a policy for motion control of an underwater vehicle via the application of model-based reinforced learning based on the use of a GP model. Reinforced learning with GPs model was shown to be a faster option than non-model standard reinforced learning techniques with less quantity of experiments and calibration required over the platform to learn a policy to control underwater vehicles. The results presented in this thesis prove that GPs are able to model autonomous underwater vehicle in an effective way by the application of multi-output GPs, allowing a series of possible applications in underwater vehicles. From all the possible applications, this thesis explored the application for post-processing of sensors data for measurement of speed-based on the difference of pressure around the head of the vehicle while the vehicle is in motion. Also, it was researched their employment as the internal model for inertial navigation algorithms such as the unscented Kalman filter for improvement in the prediction and correction of measurement. A final application of GPs for model based reinforced learning algorithms was as a reference model to allow the learning of policies for motion control for underwater vehicles. The contributions of this thesis start by establishing that GPs are capable of effectively creating a non-parametric model-based on Gaussian distribution from an underwater vehicle data. Additionally, it is established that GPs non-parametric models are a solution for the creation of transfer functions between raw sensor data as a pressure sensor array to vehicle speed for AUVs. It has been shown that GPs non-parametric models from underwater vehicles allow more robust navigation system by their integration as the internal model in an unscented Kalman filter. Also, their application to navigation to allow a faster calibration by removing the requirement of tuning of the variance matrices of state and measurement as these matrices are produced by the non-parametric model. A final contribution of this thesis is the use of GPs non-parametric models for model based reinforced learning by the search of policies for waypoint tracking and path tracking with a low quantity of trials require and overperforming common robust PID controllers