Planejamento de trajetória explorativa e informativa para monitoramento de ambientes desconhecidos com UAV: uma abordagem BO-POMDP

Abstract

Unmanned Aerial Vehicles (UAVs) have been used for several applications in monitoring complex and unknown environments. The challenge is to plan missions for the UAV where the vehicle needs to visit and explore an area and analyze it in real time to define the route to be followed. This visit involves maximizing the search area through trajectory definitions, enabling information collection to gain Knowledge about the environment and generate a map. This type of problem is known as Informative Path Planning and Autonomous Exploration. In this context, Bayesian Optimization (BO) has been used together with Gaussian Process to collect information from an unknown environment and define the trajectories. Moreover, However, for the definition of trajectories to be carried out sequentially and specific constraints to be considered, there is a need for a planner responsible for making these decisions and, in this case, the Partially Observable Markov Decision Process (POMDP) can be used. POMDP is considered a non-myopic method as it allows you to efficiently calculate the best action by checking several steps ahead. Therefore, considering these methods, the aim of this work is to develop a strategy for explorative and informative path planning in unknown environments using UAV. For this approach, a system for sequential decision making under uncertainty was developed based on the BO-POMDP approach. Tests were carried out and demonstrated that the non-Myopic approach to exploration/exploitation obtained better results compared to the Myopic and random methods. With the RMSE and Tr(P) metrics used to evaluate the models, it was inferred that the non-myopic approach obtained lower error and that uncertainty about the environment decreases throughout the algorithm iterations. Furthermore, the results demonstrated that the use of the non-myopic strategy led to a reduction in the trajectory to be covered by the UAV in a time that was equivalent to the time of the myopic method. This fact corroborates the idea that non-myopic algorithms can efficiently calculate the best action, taking shorter trajectories and, due to planning aiming for future rewards, obtaining a greater chance of finding global optimal solutions.