Search

The proposed navigation system, based on a topological representation of the world, allows the robot to robustly navigate in corridor and structured environments. This is a very practical issue in assistance applications, in which robots must perform guidance missions from room to room in environments typically structured in corridors and rooms, such as hospitals or nursing homes. Although the topological map consists of very simple and reduced information about the environment, a set of robust local navigation behaviors (the actions of the model) allow the robot to locally move in corridors, reacting to sensor information and avoiding collisions, without any previous metric information. Another important subject in robot navigation is robustness in dynamic environments. It is demonstrated that topological representations are more robust to dynamic changes of the environment (people, obstacles, doors state, etc.) because they are not modelled in the map. In this case, in which local navigation is also based on an extracted local model of the corridor, the system is quite robust to people traversing the corridor. People are another source of uncertainty in actions and observations, which is successfully treated by the probabilistic transition and observation models. Regarding doors state, the learning module adapts the probabilities to its real state, making the system more robust to this dynamic aspect of the environment. In order to improve the navigation capabilities of the proposed system, we are working on several future work lines. The first one is to enlarge the action and observation sets to navigate in more complex or generic environments. For example, to traverse large halls or unstructured areas, a "wall-following" or "trajectory-following" action would be useful. Besides, we are also working on the incorporation of new observations from new sensors, such as a compass (to discriminate the four orientations of the graph) and a wireless signal strength sensor. Enlarging the model doesn't affect the proposed global navigation algorithms. Regarding the learning system, future work is focused on automatically learning the POMDP structure from real data, making even easier the installation process. Another current research lines are the extension of localization, planning and learning probabilistic algorithms to multi-robot cooperative systems (SIMCA project) and the use of hierarchical topological models to expand the navigation system to larger structured environments.