Kooperative autonome Exploration in der Außenbereichsrobotik

Autonomous exploration is a complex task for outdoor robots which is concerned with navigation in and mapping of previously unknown outdoor areas. Motivated by the possibility of automatically exploring parts of the world that are inaccessible to or too dangerous for humans, it has to be stated that the value of robots that can handle exploration tasks is very high. Space, deep sea or polar region exploration are just a few prominent of the many examples of such systems. Complex obstacle configurations and terrain types encountered in semi- or unstructured environments present significant challenges for robotic exploration systems. Demanding complete autonomy of the operating robots increases the difficulty of the task even further. A large variety of topics, each being a separate field of research, has to be dealt with, when building a fully integrated robotic system in this context.
This thesis proposes a solution to this complex undertaking and describes an integrated system, which is able to execute the task of autonomous exploration in an unstructured outdoor environment. A cooperating team consisting of a ground and an aerial robot, AMOR and PSYCHE, is designed to tackle the problem at hand. The aerial robot acts as an external sensor platform for the ground robot, while always following the ground robot. The additional and fundamentally different angle of view on the current scene provided by the aerial robot allows for completely new approaches to solving the posed problems. The scientific innovations presented in this work are focused on the three essential research topics: localization, environment modelling, and path planning. Additionally the conception and implementation of a fully functional, integrated robotic system is addressed.
Two new approaches for solving the task of localizing robotic systems in an earth fixed coordinate system are introduced, which do not build on the well-established yet sometimes unreliable GPS technology. The first approach a probabilistic algorithm for estimating the current pose of the ground robot is developed, which makes exclusive use of data data from a dead-reckoning sensor and metric-topological map data. The second approach realizes a registration of live aerial imagery of the cooperating flying robot with a geo-registered orthophoto from a geo-database to achieve a precise localization of the ground robot by identifying its position in the live aerial image.
Environment modelling is achieved by fusing data of multiple redundant as well as complementary sensors. Actuated laserscanners are used to create a description of the geometric structure of the environment in form of a triangle mesh, which is an advancement to the common approach of using unstructured point clouds for this task. The geometric model is additionally enhanced with visual information originating from different visual sensor sources. Beside cameras with entocentric and catadipotric lenses mounted on the ground robot, an airborne camera mounted on the aerial robot is used, which constitutes an innovative evolution of the current state-of-the-art in this context. The resulting model of the environment can be used to solve problems that could not be solved by using a strictly geometric model previously. Distinguishing between traversable and non-traversable planar surfaces is only one of many examples that underline the benefit of a visually enriched geometric model for outdoor robotics.
Path planning in the context of an exploring autonomous robot stands for the planning of collision-free paths that enable the robot to escape from the known environment and lead it into the unknown. Unfortunately algorithms focusing on exploration are underrepresented among the currently available path planning algorithms, which often focus on the classical concept of start-goal-planning. Due to this shortcoming a novel method for local explorative path planning based on the concept of Rapidly Exploring Random Trees is proposed in this thesis. It focuses on the efficient determination of all so-called exploration paths. The new method of local explorative path planning uses classical geometrical features as well as visual features to guide the path planning process.
The developed algorithms have been experimentally evaluated in several real-world situations with the fully autonomous robot team described in this thesis.