Mobile robots have to find their way around unfamiliar environments in order to help people in their everyday lives. This requires complex algorithms and artificial intelligence in the systems. Since October 2018, Prof. Dr. Ute Bauer-Wersing from the Frankfurt University of Applied Sciences (Frankfurt UAS) has been researching the visual self-localization and navigation of mobile robots in a collaborative project with the Honda Research Institute Europe. Based on the research results to date, the project 'Unsupervised Learning of Hierarchical Features for Visual Self-Localization and Navigation in Seasonally Changing Outdoor Environments' has now been extended by one year until the end of September 2022. This increases the volume of third-party funding, which the 'Machine Learning for Intelligent Systems' research group has successfully completed in several projects with the cooperation partner in recent years, to more than one million euros.

From the lab to everyday life

Prof. Dr. Ute Bauer-Wersing, Frankfurt UAS

© Private

According to Prof. Dr. Bauer-Wersing, the possibility of machines developing intelligent behavior with the help of learning algorithms promises great potential for various fields of application. Systems of this kind could be used for vehicles that navigate and move largely autonomously in road traffic, for flying drones that transport parcels independently, or for mobile service and household robots that support people at work and in everyday life. Intelligent systems are already showing promising capabilities under laboratory conditions. However, the major challenge is to ensure that the systems also function reliably in complex new environments and can adapt dynamically to any changes that occur.

This is where the cooperation project of the 'Machine Learning for Intelligent Systems' research group headed by Prof. Dr. Bauer-Wersing and the Honda Research Institute Europe comes in. "The ability to localize itself in a complex environment, e.g. in an apartment or a garden, with the help of camera images and to navigate reliably to a specific position is one of the basic learning tasks that an autonomous robot must master," explains Prof. Dr. Bauer-Wersing. "This task is particularly difficult in outdoor areas, where weather conditions and changing vegetation periods constantly and drastically change the appearance of the surroundings. As humans and animals solve this task without any problems, our approach is based on the brain's positioning system."

Localization based on camera images

The research team is developing learning algorithms that replicate cognitive maps in the form of so-called 'place cells', which are among the most important components of the brain's internal GPS. The biologically-inspired approach offers several advantages: Previous research has shown that precise localization down to a few centimetres is possible based on camera images. The representations learned by the machine are robust in the face of many dynamic changes. The learned characteristics also allow the robot to move purposefully in the environment and the routes chosen by the robot are similar to those that humans and animals would choose.

"The project results to date have shown that the approach is both very productive in terms of gaining scientific knowledge and very viable with a view to practical application in service and household robots of the generation after next," says Prof. Dr. Bauer-Wersing. Due to the successful outcome, the third-party funded project has now been extended for a further year until September 2022.