Placed Perales, Julio Alberto
Castellanos Gómez, José Ángel (dir.)
Universidad de Zaragoza,
2024
Resumen: Mobile robotics has undergone major advances in recent years, with the ultimate goal of deploying fully autonomous robots in the real world capable of performing complex missions, such as disaster relief search and rescue, deep-sea and planetary exploration, and service robotics. However, in order to operate effectively in such environments, robots must first build a model of the environment over which they can reason. Simultaneous localization and mapping (SLAM) allows a robot to build such a model (usually in the form of a map) while simultaneously determining its own location on it. Yet, these are passive approaches in which the robot either follows a predetermined trajectory or is tele-operated while forming the model of the environment; hence limiting its autonomy.
Active SLAM incorporates the navigation aspect into the problem, therefore playing a fundamental role in the deployment of autonomous agents in the real world. Compared to SLAM, the objective has now shifted to having the robot decide its own future movements while performing SLAM in order to build the best possible representation of the environment. To do this, the robot must weigh the costs and benefits (i.e., utility) of executing different actions and find those that reduce the uncertainty of its localization and the map representation.
This thesis aims to advance the frontiers of active SLAM, by presenting new methods for effective, fast, and optimal decision-making. The main contributions of this thesis can be divided into four main parts, each of which is associated with a major challenge in the field.
Active SLAM has been studied in different forms across multiple communities, resulting in numerous approaches based on different concepts and theories that have made the field thrive, but also created a lack of unification that hinders the problem understanding, and a disconnect between lines of research that could mutually benefit from each other. By presenting a comprehensive survey, we address the need for a unified problem formulation and provide a guide for both researchers and practitioners. We offer a new perspective of the problem and outline a number of open challenges and promising research directions.
Regardless of the theoretical background or the approach taken, the evaluation of the utility of performing a given set of actions prevails in most active SLAM methods; and this boils down to estimating the uncertainty of future states. This is a complex and time-consuming step that often constitutes the main bottleneck. To alleviate this, we explore the opportunities that spectral graph theory offers and how it can be leveraged to speed up uncertainty quantification during active graph-SLAM. We derive a theoretical relationship between the well-established optimality criteria and the graph connectivity indices that allows uncertainty quantification in just a fraction of the time required by classical methods. This lays the foundation for topological active SLAM, or spectral active SLAM. Besides, we demonstrate the usefulness of these novel techniques by presenting three applications: two open-source end-to-end systems that make optimal decisions online using the graph topology, and a novel stopping criterion. We show that these methods in particular, and topological utility functions in general, yield decisions equivalent to using classical utility functions in a fraction of the time.
Advances in deep learning have also opened a new avenue for rapid decision-making. They offer the opportunity to move the costly uncertainty quantification to an offline training phase, reducing real-time operation to a forward pass on the network. We present a novel end-to-end approach to active SLAM based on uncertainty-aware deep reinforcement learning. Unlike most existing approaches, we go beyond neural obstacle avoidance and train agents capable of making uncertainty-informed decisions in real-time by embedding classical utility functions in the reward design. Thus, we provide a link between estimation-theoretic and data-driven approaches. We demonstrate the feasibility of uncertainty-aware learning and show that uncertainty quantification can be learned during active SLAM.
Finally, we investigate reasoning over high-level representations in active SLAM. Humans perceive and represent the environment in a very different way than traditional robots have traditionally done. Of course, we have a geometric level of understanding of it, but reasoning usually goes beyond that: semantics, abstract high-level entities and the relationships between them are crucial. Very recent work has gone into incorporating these abstract concepts into models of the environment, thus endowing robots with spatial perception. As with SLAM, this has raised the question of how to build such models autonomously by reasoning about their uncertainty, that is, active spatial perception. We present a general formulation to quantify uncertainty over these novel representations and a first approach to tackle this challenge, leveraging the structure and hierarchy of the model.
In summary, this thesis addresses several of the current major challenges in the field of active SLAM. We contribute a comprehensive survey of the problem and solutions to achieve fast decision-making and reasoning over high-level representations. In addition, we make a great effort towards reproducible and comparable research by open-sourcing the code for all of the aforementioned methods.
Resumen (otro idioma):
Active SLAM incorporates the navigation aspect into the problem, therefore playing a fundamental role in the deployment of autonomous agents in the real world. Compared to SLAM, the objective has now shifted to having the robot decide its own future movements while performing SLAM in order to build the best possible representation of the environment. To do this, the robot must weigh the costs and benefits (i.e., utility) of executing different actions and find those that reduce the uncertainty of its localization and the map representation.
This thesis aims to advance the frontiers of active SLAM, by presenting new methods for effective, fast, and optimal decision-making. The main contributions of this thesis can be divided into four main parts, each of which is associated with a major challenge in the field.
Active SLAM has been studied in different forms across multiple communities, resulting in numerous approaches based on different concepts and theories that have made the field thrive, but also created a lack of unification that hinders the problem understanding, and a disconnect between lines of research that could mutually benefit from each other. By presenting a comprehensive survey, we address the need for a unified problem formulation and provide a guide for both researchers and practitioners. We offer a new perspective of the problem and outline a number of open challenges and promising research directions.
Regardless of the theoretical background or the approach taken, the evaluation of the utility of performing a given set of actions prevails in most active SLAM methods; and this boils down to estimating the uncertainty of future states. This is a complex and time-consuming step that often constitutes the main bottleneck. To alleviate this, we explore the opportunities that spectral graph theory offers and how it can be leveraged to speed up uncertainty quantification during active graph-SLAM. We derive a theoretical relationship between the well-established optimality criteria and the graph connectivity indices that allows uncertainty quantification in just a fraction of the time required by classical methods. This lays the foundation for topological active SLAM, or spectral active SLAM. Besides, we demonstrate the usefulness of these novel techniques by presenting three applications: two open-source end-to-end systems that make optimal decisions online using the graph topology, and a novel stopping criterion. We show that these methods in particular, and topological utility functions in general, yield decisions equivalent to using classical utility functions in a fraction of the time.
Advances in deep learning have also opened a new avenue for rapid decision-making. They offer the opportunity to move the costly uncertainty quantification to an offline training phase, reducing real-time operation to a forward pass on the network. We present a novel end-to-end approach to active SLAM based on uncertainty-aware deep reinforcement learning. Unlike most existing approaches, we go beyond neural obstacle avoidance and train agents capable of making uncertainty-informed decisions in real-time by embedding classical utility functions in the reward design. Thus, we provide a link between estimation-theoretic and data-driven approaches. We demonstrate the feasibility of uncertainty-aware learning and show that uncertainty quantification can be learned during active SLAM.
Finally, we investigate reasoning over high-level representations in active SLAM. Humans perceive and represent the environment in a very different way than traditional robots have traditionally done. Of course, we have a geometric level of understanding of it, but reasoning usually goes beyond that: semantics, abstract high-level entities and the relationships between them are crucial. Very recent work has gone into incorporating these abstract concepts into models of the environment, thus endowing robots with spatial perception. As with SLAM, this has raised the question of how to build such models autonomously by reasoning about their uncertainty, that is, active spatial perception. We present a general formulation to quantify uncertainty over these novel representations and a first approach to tackle this challenge, leveraging the structure and hierarchy of the model.
In summary, this thesis addresses several of the current major challenges in the field of active SLAM. We contribute a comprehensive survey of the problem and solutions to achieve fast decision-making and reasoning over high-level representations. In addition, we make a great effort towards reproducible and comparable research by open-sourcing the code for all of the aforementioned methods.
Resumen (otro idioma):
Pal. clave: robótica
Titulación: Programa de Doctorado en Ingeniería de Sistemas e Informática
Plan(es): Plan 512
Área de conocimiento: Ingeniería y Arquitectura
Nota: Presentado: 08 03 2024
Nota: Tesis-Univ. Zaragoza, , 2024
Aportación del TFG/M a la Sostenibilidad: Garantizar una educación de calidad inclusiva y equitativa, y promover las oportunidades de aprendizaje permanente para todos. Fomentar el crecimiento económico sostenido, inclusivo y sostenible, el empleo pleno y productivo, y el trabajo decente para todos. Desarrollar infraestructuras resilientes, promover la industrialización inclusiva y sostenible, y fomentar la innovación.
Enlace permanente:
Visitas y descargas
Registro creado el 2024-06-20, última modificación el 2024-06-20
