LIP6 2002/001
- Thesis
Cartographie et estimation globale de la position pour un robot mobile autonome - D. Filliat
- 182 pages - 12/21/2001- document en - http://www.lip6.fr/lip6/reports/2002/lip6.2002.001.pdf - 3,320 Ko
- Contact : david.filliat (at) nulletca.fr
- Ancien Thème : OASIS
- Keywords : navigation, animat, global localization, map-learning, path-planning
- Publisher : Ghislaine.Mary (at) nulllip6.fr
Managing the movements of an autonomous mobile robot in its environment is a problem that has been tackled since the early integration of artificial intelligence and robotics. However, this problem remains difficult and no general solution has been devised. Among existing navigation strategies, we will focus on those that use a map to represent the spatial layout of the environment and that allow to plan movements toward distant goals. Map-building and self-positioning within these maps are two sub-problems that have been solved independently. However, solving these sub-problems simultaneously is still a difficult task.
In particular, position estimation decoupled from map-building can efficiently be realized by probabilistic models based on Partially Observable Markov Decision Processes. These models however cannot generally be used to build the map they use on-line.
We have designed a navigation model inspired from these localization methods that is able to build a map on-line using only relatively imprecise sensors. The capacity to integrate various information sources afforded by this method makes it possible to compensate the low quality of a single perception. Moreover, active perception procedures allow to efficiently use available sensors with regard to the current context. Thus, our model allows a robust position estimation along with simultaneous map-building with few hypotheses on the environment. Its capacities have been demonstrated in simulation and on a real robot. In particular, we have shown that it allows a robot to localize itself correctly as soon at it is introduced in a new environment, to rapidly relocalize itself if it is passively transported from one place to another, and to plan a reliable trajectory to a distant goal.
In particular, position estimation decoupled from map-building can efficiently be realized by probabilistic models based on Partially Observable Markov Decision Processes. These models however cannot generally be used to build the map they use on-line.
We have designed a navigation model inspired from these localization methods that is able to build a map on-line using only relatively imprecise sensors. The capacity to integrate various information sources afforded by this method makes it possible to compensate the low quality of a single perception. Moreover, active perception procedures allow to efficiently use available sensors with regard to the current context. Thus, our model allows a robust position estimation along with simultaneous map-building with few hypotheses on the environment. Its capacities have been demonstrated in simulation and on a real robot. In particular, we have shown that it allows a robot to localize itself correctly as soon at it is introduced in a new environment, to rapidly relocalize itself if it is passively transported from one place to another, and to plan a reliable trajectory to a distant goal.