| Imprimer |
| Autonomous exploration of an unknown 3D environment (Exploration autonome d'un environnement 3D inconnu) Noël, Thibault - (2025-01-24) / Université de Rennes Autonomous exploration of an unknown 3D environment
| |||
Langue : Anglais Directeur(s) de thèse: Chaumette, François; Marchand, Éric Discipline : Automatique, productique et robotique Laboratoire : INRIA-RENNES Ecole Doctorale : MATISSE Classification : Sciences de l'ingénieur, Informatique Mots-clés : Planification basée capteurs, Robotique mobile, Exploration
| |||
Résumé : Ce manuscrit présente les méthodes développées pour permettre l’exploration autonome d’un environnement 3D inconnu par un robot mobile. L’exploration est une tâche transverse, regroupant divers sous-problèmes ; ici, nous nous intéressons en particulier à la planification de trajectoires, essentielle pour la navigation autonome du robot. Nous proposons en particulier deux méthodes nouvelles de construction de graphes de navigation, utilisées ensuite comme base pour l’élaboration de stratégies d’exploration plus haut-niveau. Deux approches principales sont présentées, reflétant les deux grands courants de la littérature sur le sujet : l’exploration guidée par les frontières et l’exploration par maximisation de l’information. Les méthodes proposées sont évaluées en simulation et dans divers environnements réels. Abstract : This manuscript investigates the problem of robotic autonomous exploration in 3D environments. It is a broad and transversal research question involving various robotics tasks such as simultaneous localization and mapping, motion planning, and control. From a high-level standpoint, exploration can be seen as the maximization of the information gain about the environment, often combined with a minimization of energy expenditure. Considering the standard action-perception loop, exploration presents a particularly strong coupling between past sensor observations and future actions, as the progressive discovery of the environment dynamically dictates the future optimal sensor path. This coupling is even more pronounced for directional sensors, e.g. cameras, for which the robot state can drastically impact the sensors observations. Deriving high-quality plans from the current state of the environment is thus essential, but remains challenging due to the high uncertainties and the large size of the workspace. In this work, we chose to focus on those planning aspects, exploring and comparing various combinations of environment representations, processing methods and planning strategies to improve the exploration performance of the robot. More precisely, we relied on standard mapping algorithms to provide a geometrical representation of the environment. The core of our work has been to exploit this representation to derive efficient navigation roadmaps and exploration strategies, taking into account the safety requirements imposed by the unknown environment while ensuring high-quality observations even for field-of-view-constrained sensors. Our results are evaluated extensively in realistic simulation environments. We also deploy our strategies in various real-world experiments, demonstrating their usability with a ground mobile robot and various sensor configurations. | |||