Mobile robot navigation in completely unprepared indoor environment remains a quite open challenge when using mostly computer vision. We aim at developing topological approaches to navigation that could cope, with the help of a user, with these challenges.
We focus on applications for personal robotics, in indoor unprepared environments, that could be easily used by non-specialist users. We mainly study vision-based navigation with different a priori :
Interactive Learning : Taking inspiration from the talking robot project, we aim at developing methods that learn by asking « well chosen » questions to the user in a « game » to achieve a task. This allows to ensure a proper mapping from the robot perceptions to the user objectives.
Active perception : We work on methods that search for informative images. This feature is compulsory for the navigation to be robust to « weird » robot position (under table, facing wall …) that will occur often in real applications.
Vision only : We develop systems with as less as possible dependence on odometer or inertia so as to produce algorithms robust to any user manipulation of the robot, such as robot displacement to distant locations.
The system we developed is strongly inspired by the "bag of visual words" method used for image classification that use a representation of images as sets of unordered local features (see this good course on the subject).
We adapted this method to the robotic context, in particular to make it fully incremental at all stages, from the dictionary construction to the learning method for the task. We also use several feature spaces in combination to enhance robustness of the approach in the different environments the robot will encounter :
QUALITATIVE LOCALIZATION AND MAPPING
We currently have several applications of this framework. The first is a qualitative localization and mapping method that learns to recognize the room the robot is in (See our article in ICRA 07 and the slides of the presentation for details). We validated this method in real environments using aibo robots, with URBI for development.
Example of environment in which the qualitative localization system was validated
The approach was also validated on the challenging INDECS database. The next figure shows the results obtained with correct room recognition in green and errors in red. Learning was performed incrementally only when errors where committed. The global success rate is around 80%, with a steady performance of more than 90% after the initial environment discovery.
We also developed a visual homing method that learns a route to a goal and is able to replay it using only vision information. This video illustrate the method :
||Interactive learning of visual topological navigation. Proceedings of the 2008 IEEE International Conference on Intelligent Robots and Systems (IROS 2008).
||A visual bag of words method for interactive qualitative localization and mapping. Proceedings of the International Conference on Robotics and Automation (ICRA).
>> David FILLIAT is the project leader.
>> Jose Luis Susa, Roland Goffettre and Florian Vichot worked on this project during their internship.