Repository of DUTH

Progressively Integrating Autonomy and Intelligence in Robotic Platforms (Doctoral thesis)

Ράπτης, Εμμανουήλ/ Raptis, Emmanuel

In this thesis, we focus on the integration of robot autonomy and intelligence for UAV-based environmental mapping and robot task planning, moving step-by-step from fundamental autonomous motion planning strategies to the integration of advanced intelligence, decisionmaking and logical reasoning. The structure of this thesis is divided into three main pillars, each progressively building upon the understanding and reasoning of the previous while gradually incorporating more advanced layers of autonomy and intelligence into robotics. The first pillar, separated in two parts, deals with the problem of determining an optimal path that passes through all the points of a known (part 1) or partially known (part 2) area while avoiding any obstacles defined within the operational environment. This problem, which is usually referred to as Coverage Path Planning (CPP), has been proven to be NPhard, with the computational time required to solve it increasing drastically as it gets more complicated (e.g. involving convex or non-convex shaped regions/ obstacles, multiple robots, etc.). As to the first part, Chapter 3 addresses the CPP for a single UAV, by proposing a novel methodology capable of producing optimal, collision-free paths in approximately polynomial time. This methodology serves as the foundational layer for robot autonomy and ensures that a single robot can perform efficient and safe navigation within a known environment. The second part expands the exploration of autonomy to real-time adaptive decision-making through informative path planning. Unlike the primary offline planning for a single UAV, this part addresses the challenge of dynamically adjusting a robot’s behaviour by processing path conditions in real-time as the agent moves toward its goal. In our effort to introduce the added layer of intelligence, we observed that in the initial layer of autonomy, the “fear of missing out” data can lead to uniform, conservative scanning policies over the whole area. Consequently, employing a non-uniform scanning approach can mitigate the expenditure of time in areas with minimal or negligible real value, while ensuring enhanced precision in information-dense regions. Turning to the available informative path planning methodologies, we discern that certain methods entail intensive computational requirements, while others necessitate training on an ideal world simulator. In Chapter 4, we propose an active sensing coverage and exploration approach, named OverFOMO, standing for Overcome the Fear of Missing Out, that regulates the speed of the UAV based on the value and confidence level of the information detected. This non-uniform scanning approach allows robots to prioritize areas rich in valuable information, minimizing time spent in less relevant regions. Having established two robust methods for single-UAV CPP, in the second pillar, we extend this autonomy to multi-agent systems, focusing on cooperative behaviour, where UAVs work in swarms. In the multi-UAV setup, we aim not only to generate collision-free paths but also to optimize energy consumption and reduce mission duration. From its nature, the utilization of multiple robots can introduce significant benefits in terms of time and energy needed to cover an area. However, to fully maximize the potential benefits of unmanned vehicles, a multi-robot CPP approach must integrate characteristics that promote overall efficiency in terms of time and energy usage. Chapter 5 introduces cooperative decision-making through an optimization scheme that maximizes the efficiency of multi-robot missions incorporating by design the knowledge of the environment and operational constraints, i.e. maximum operational time and remaining battery. The addition of intelligence in this phase emerges from the optimization of various operational parameters, such as path lengths, the number of turns, re-visitations, and launch positions. This progression from single-agent to multi-agent autonomy enables increasingly intelligent robotic systems, as we move toward optimizing group behaviour in complex environments. Finally, the third pillar moves one step further, integrating autonomy and intelligence into a unified system for any robotic platform. In this phase, we shift from task-specific intelligence to a broader, more flexible form of intelligence, referred to as Intelligence Quotient (IQ). IQ in robotic platforms typically refers to the cognitive abilities of a robotic system and its level of sophistication in terms of problem-solving, learning, adaptability, and decision-making. Yet, robot intelligence goes beyond cognitive abilities; it demands a deep understanding of commonsense and the ability to interact with users, translating human natural language into a logical sequence of robotic actions that are both physically feasible and coherent in the real world. In Chapter 6, we introduce RobotIQ, a novel framework that empowers mobile robots with human-level planning capabilities, enabling seamless communication via natural language instructions with any Large Language Model (LLM). Built upon the Robotic Operating System (ROS), RobotIQ offers a modular crafted robot library suite of API-wise control functions focusing on navigation, manipulation, perception, localization, and human-robot interaction. It simplifies the integration of intelligence (IQ) across various robotic platforms, including ground, aerial, and underwater systems, as well as robotic arms. The effectiveness of the developed methodologies across all pillars was validated through simulated and real-world experiments. Starting with fundamental path planning towards a fully autonomous deployment of a single robot for coverage in known environments, this thesis moves towards real-time adaptive decision-making for exploration in partially known environments, extends to the coordination of multi-robots while incorporating energy-aware characteristics and ultimately culminates in a unified, intelligent framework for empowering mobile robots with human-level reasoning. The developed solutions are modular, and can be tailored to a wide range of applications.