Résumé : |
(auteur) This thesis addresses the problem of autonomous navigation with ground robots in complex environments, which may be characterized as nonplanar and nonstatic. The goal of the presented research is to enable reliable navigation over large distances in generic indoor and outdoor environments, independent of external localization sources such as a global positioning system (GPS). Focusing on these challenges, algorithms for all building blocks of autonomous navigation—localization, mapping, terrain assessment, motion planning, and motion control—are developed, implemented, integrated, and finally evaluated in extensive field experiments. Sensor-based perception of the environment is a basic requirement for localization and mapping. We propose to use a high-frequency three-dimensional (3D) laser scanner as the main exteroceptive sensor. The advantages of this technology lie in the high density and accuracy of the provided measurements, and their independence of lighting and weather conditions. We develop a highly scalable system for six-dimensional (6D) localization and 3D mapping based on iterative closest point (ICP) matching. A topological/metric map representation, where metric information is kept in spatially constrained local submaps representing vertices in a graph, allows to build consistent large-scale maps without requiring global optimization. Long-term application in dynamic and changing environments is enabled by integrating methods for identifying dynamic objects in the scene and for continuously updating existing submaps. Planning feasible and safe motions for a robotic vehicle requires distinguishing obstacles from traversable terrain. We develop two different algorithms for terrain assessment. The first method is targeted at real-time obstacle detection in the vicinity of the robot. Assuming locally planar terrain, a grid-based obstacle map is built by analyzing the raw laser scans. The second approach is based on dense point cloud maps (which can be obtained from the ICP mapping system) and suitable for planar and nonplanar environments. The algorithm computes the geometry and the traversability of the terrain “on demand” at specific query locations, avoiding any artificial discretization or explicit surface reconstruction. The desired terrain characteristics are estimated based on statistics on the local distribution of map points. Given a specific navigation task, motion planning can be defined as the problem of reasoning about how to act based on the knowledge about the environment. This thesis addresses both local obstacle avoidance and global planning over large distances. Our approach to local planning consists of computing a set of candidate trajectories, which are shaped around nearby obstacles or along a given reference path, and enforced to satisfy the robot’s kinematic constraints. The optimal local trajectory is chosen by evaluating the motion alternatives in terms of guidance towards the goal and traversability of the underlying terrain. For global motion planning, we develop an algorithm embedding the proposed point-cloud-based terrain assessment method, which allows trajectories to be directly planned on 3D point cloud maps. The approach is designed to be suitable for generic nonplanar environments, including rough outdoor terrain, multi-level facilities, and more complex geometries. Piecewise continuous trajectories are computed in the full 6D space of robot poses, while strictly considering the vehicle’s kinematic and dynamic constraints. We apply sampling-based planning algorithms to generate an initial trajectory connecting the desired start and goal poses. Subsequently, the trajectory is locally optimized according to a generic cost function, which may include path length, path curvature, and roughness of the traversed terrain. While enforcing the hard constraints to remain satisfied (terrain contact, traversability, kinodynamic feasibility), the trajectory is iteratively deformed until a local minimum of the cost function is reached. We develop two complete systems for autonomous navigation, integrating these approaches. Combining the ICP-based localization and mapping framework with local obstacle detection and local motion planning, we implement a framework for autonomous route following, commonly referred to as teach and repeat (T&R). After a manually controlled teach run, where a graph of local submaps is built, the robot is able to automatically repeat the learned route, using the recorded maps for localization. Unlike classical T&R systems, our framework is suitable for application in dynamic environments, where the integrated obstacle avoidance scheme allows to detect and circumnavigate obstacles appearing on the reference path. In addition to the T&R approach, we present a second navigation system, integrating the point-cloud-based terrain assessment and global planning algorithms with ICP-based localization and mapping. Given a graph of point cloud maps—typically recorded in a manually controlled survey run—the framework enables navigation within the mapped area without being restricted to known routes. Motion control is implemented by a trajectory tracking controller with integrated real-time collision checking. Together with continuous map updates and frequent replanning of the global trajectory, these techniques enable autonomous navigation in nonplanar, nonstatic environments. Finally, we describe the characteristics of the mobile robot ARTOR, which was set up for the purpose of testing and evaluating the developed algorithms under realistic conditions. ARTOR consists of a six-wheeled, electrically powered base vehicle equipped with sensors, computers, and communication gear. The proposed autonomous navigation algorithms were integrated on the robot and tested in extensive field experiments, demonstrating reliable, GPS-independent navigation over large distances and under greatly varying environmental conditions, in unstructured off-road terrain, multi-level environments, and dynamic urban areas. |