Task Planning for fully autonomous underwater operations with AI-based environment perceiving and modeling

Autonomous Underwater Vehicles (AUVs) have become fundamental tools for marine scientists and industries to explore and monitor underwater areas. Indeed, non-predictable environmental conditions and sensor acquisitions make the design of AUV surveys challenging even for expert operators. Multiple attempts are required, and the collected data quality is not guaranteed: the AUV usually passively stores the sensor acquisitions that are then analyzed offline after its recovery by human operators in charge of identifying and localizing the so-called Objects of Potential Interest (OPIs). When it comes to acoustic images, the lack of features and low resolution make this task even more challenging. As a consequence of these statements, the marine community has sought robots able to meaningfully perceive and model the surroundings and autonomously conduct the assigned task, which are the foundations of fully autonomous vehicles. In order to provide a compact and lightweight AUV with the capability to gather knowledge of the subsea scenario, an Automatic Target Recognition (ATR) strategy based on modern Convolutional Neural Networks (CNNs) for onboard online applications has been developed and tested. More specifically, the resulting ATR methodology has been used to detect potential targets of interest in Forward-Looking Sonar (FLS) imagery. Such ATR architecture has been incorporated with two developed world modeling procedures aimed at creating a semantically enriched environment representation comprising of 3D-localized and labeled objects of interest. The extended Probabilistic Multiple Hypothesis Anchoring (PMHA) methodology has been tested over a multi-vehicle scenario for OPIs placed in the seabed. The Probabilistic Particle Filter Anchoring (PPFA) is proposed with the aim of addressing the ambiguity space of the employed perception device (e.g., range of detected objects with cameras and elevation information with imaging FLS) and providing a probabilistic semantic 3D representation of the underwater world in completely unknown scenarios. Finally, the novel Multi-Hypothesis Task Planning (MH-TP) architecture has been designed to integrate temporal Artificial Intelligence based (AI) planning with the aforementioned semantic world modeling techniques toward fully autonomous AUV inspections in unknown environments. The proposed architecture extends the standard AI planning methodologies relying on the Problem Domain Definition Language (PDDL) by dynamically varying the PDDL problem with the probabilistic, semantically enriched objects that the World Modeling (WM) itself evaluates. The solutions proposed in this thesis have been firstly validated with realistic simulations made by means of the Unmanned Underwater Vehicles (UUVs) Simulator, where a dynamic model of FeelHippo AUV, developed by the Department of Industrial Engineering of the University of Florence, was implemented alongside the requested perception and payload devices. Moreover, the ATR strategy, the PMHA and PPFA frameworks have been tested with on-field gathered data.