Underwater Glider Path Planning Under Partially Observable Markov Decision Processes Framework

Underwater gliders are one type of underwater platforms for oceanographic observations.They are becoming popular autonomous underwater vehicles and have attracted researchers' attention because of the advantages of low cost,long endurance,wide application,and the potential to be extended for large-scale observations.Path planning is one important aspect of the underwater glider's autonomy.On one hand,autonomous obstacle avoidance as one essential quality is the premise to ensure safe navigation,especially when the glider is observing in a chaotic environment,such as the coast,static and dynamic underwater obstacles pose a threat to its safety.On the other hand,intelligent path planning has gradually proposed to achieve efficient environmental observation because the energy that the underwater glider can carry is very limited.This thesis focuses on path planning of the underwater glider based on Partially Observable Markov Decision Processes(POMDPs).A three-dimensional obstacle avoidance model and an adaptive sampling model are established under the POMDP framework.The main contributions are as follows:(1)When underwater gliders are sailing in the water,their motion is easily affected by the dynamic marine environment.It is also difficult to estimate their positions due to the uncertainties induced by the actuation mechanism.These uncertainties would eventually cause difference between theoretical and actual paths.In this thesis,a POMDP model is built based on the fundamental principle and performance of underwater gliders.In the model,the uncertain effects of the marine environment and the actuator on the state of underwater gliders are considered.A state transition function is also constructed to realize updates of probability distributions.(2)In the process of navigation,the underwater gliders are not able to detect the entire environment.They only have access to local environmental information through sensors.In the absence of all obstacle information,In this thesis,a path planning approach to obstacle avoidance is presented under the POMDP framework for underwater gliders.A grid map is established to present the probability of potential obstacles and is updated by particle filtering.(3)Autonomous path planning and efficient ocean observation capability are important embodiment of intelligence that underwater gliders display.In this thesis,an efficient adaptive path planning approach is presented based on the POMDP model.The glider's current state is combined in the model to maximize sampling information,so as to enable an efficient adaptive sampling.Estimation of a temperature field is presented as a case study.A Gaussian process regression(GPR)model with strong robustness is introduced to predict the temperature field.Prediction is conducted according to current available data and the relation between the data.The accuracy of the GPR model is verified by data from sea trials.