Reach that goal

Autonomous rover navigation is the original and main objective of the project, to which most of the work has been devoted up to now. It consists in endowing the rover with the capacity to autonomously execute the task "Goto Goal", the goal being up to several kilometers far away in an initally poorly known environment. For that purpose, the rover must have the ability to perceive and model the environment, to localize itself, and to generate and execute motions. All these processes must be integrated and actively controlled within a decisional real time architecture [Chatila 1995]: autonomous navigation is a canonical instance of machine intelligence, on the basis of which higher level tasks, such as exploration, reconoissance or surveillance can be easily achieved.

General approach : adaptive navigation

The following images show three kind of terrains in our experimental site, and the corresponding digital elevation maps. Would you drive your rover the same way in each of these cases ?


For efficiency purposes, it is highly desirable that the rover adapts its activities to the kind of terrain traversed. Indeed, crossing a wide open area can be achieved on the basis of a simple reactive "avoid obstacles" loop, whereas traversing a rough and dangerous area requires slowest speeds, finer terrain modeling and finer motion execution control for instance. The rover must therefore be able to choose among a set of various navigation modes [Lacroix 1994, Chatila 1997]. The definition and the number of these modes depend on the rover mechanical structure and on the kind of terrains he might encounter during its missions.

A bunch of algorithms

A consequence of this adaptive approach is that each of the main functionalities required by navigation i.e. environment perception, environment modeling, localization, motion generation and motion execution, can be achieved in various ways, depending on the current active motion mode. A simple obstacle/free binary model of the environment is sufficient to navigation in flat, open areas, whereas a fine digital elevation map is required to cross a rough area for instance. Similarly, motions can be either generated by simple reactive methods, or planned on the basis of the mechanical capacities of the rover chassis in rough areas. As for localization, it is a so important functionality that it requires the integration of various algorithms, from inertial navigation to place recognition. All these algorithms are integrated within a modular, evolutive architecture, and controlled according the context and some pre-defined navigation strategies [Lacroix 2002].