A Behavioral Approach to Indoor Autonomous Navigation with an Affordance-based Plan execution
Loading...
Date
2025
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
La mayoríaa de los sistemas de navegación autónoma utilizados actualmente en robots móviles, basan su funcionamiento en la construcción de representaciones de tipo métrica o topológica del ambiente. Desafortunadamente, este tipo de información no es robusta a cambios estructurales del entorno, y su capacidad de generalización se ve limitada al requerir de la construcción de una representación particular para cada ambiente visitado. Para superar las limitaciones antes mencionadas, proponemos un modelo de navegación autónoma capaz de representar ambientes de interior a través de: i) los distintos espacios arquitectónicos que son significativos para los seres humanos, tales como oficinas, corredores y halls, y ii) los comportamientos necesarios para transitar entre dichos espacios, tales como “salir de oficina”, “seguir corredor” o “cruzar hall”. Con este modelobuscamos alcanzar un alto nivel de abstracción, que haga al sistema menos propenso a errores causados por discrepancias entre la representación interna y el entorno real.Las principales contribuciones de este trabajo son: i) un esquema de navegación enfocado en la selección de acciones de alto nivel que es menos propenso a errores de localización, ii) una representación compacta y semánticamente rica del ambiente que facilita el planeamiento de rutas e interacción con seres humanos, y iii) un método para ejecución de planes de navegación basada en la identificación de affordances relevantes para navegación, que permite prescindir de una representación explícita del ambiente. Al evaluar el rendimiento de nuestra solución en ambientes vistos en entrenamiento, obtenemos que nuestro modelo logra superar la tasa de éxito de modelos alternativos hasta en un 43%. Por otro lado, nuestro modelo presenta resultados prometedores para ambientes no vistos, superando hasta en un 26.3% el rendimiento de modelos alternativos.
Most of the current robotic autonomous navigation systems base their operation on metric or topological representations of the environment. Unfortunately, this type of information is not robust to structural changes, and its generality is limited because it requires constructing a particular representation for each visited environment. To overcome the limitations above, we propose an autonomous navigation model capable of representing indoor environments through i) the spaces that are significant to humans, such as offices, corridors, and halls, and ii) the behaviors required to move between the spaces mentioned above, such as “leaving office”, “following corridor” or “crossinghall”. With this type of representation, we aim to achieve a high level of abstraction and avoid problems caused by geometric discrepancies between the internal representation and the real environment.The main contributions of this work are: i) a navigation scheme centered in high-level action selection that is less prone to localization errors, ii) a compact and semantically rich representation that facilitates path planning and interaction with humans and iii) a method that poses the execution of a navigation plan as the identification of relevant navigation affordances without relying on an explicit representation of the environment. Our experiments suggest that our approach performs well in previously seen environments compared to baselines, exceeding alternative approaches’ success rate by up to 43%. Additionally, our approach shows promise at generalizing to previously unseenenvironments, exceeding the success rate of alternative approaches by up to 26.3%..
Most of the current robotic autonomous navigation systems base their operation on metric or topological representations of the environment. Unfortunately, this type of information is not robust to structural changes, and its generality is limited because it requires constructing a particular representation for each visited environment. To overcome the limitations above, we propose an autonomous navigation model capable of representing indoor environments through i) the spaces that are significant to humans, such as offices, corridors, and halls, and ii) the behaviors required to move between the spaces mentioned above, such as “leaving office”, “following corridor” or “crossinghall”. With this type of representation, we aim to achieve a high level of abstraction and avoid problems caused by geometric discrepancies between the internal representation and the real environment.The main contributions of this work are: i) a navigation scheme centered in high-level action selection that is less prone to localization errors, ii) a compact and semantically rich representation that facilitates path planning and interaction with humans and iii) a method that poses the execution of a navigation plan as the identification of relevant navigation affordances without relying on an explicit representation of the environment. Our experiments suggest that our approach performs well in previously seen environments compared to baselines, exceeding alternative approaches’ success rate by up to 43%. Additionally, our approach shows promise at generalizing to previously unseenenvironments, exceeding the success rate of alternative approaches by up to 26.3%..
Description
Tesis (Doctor in Engineering Sciences)--Pontificia Universidad Católica de Chile, 2025
Keywords
Navegación basada en comportamientos, Affordance, Aprendizaje por imitación supervisado, Behavior-based navigation, Supervised imitation learning