The aim of the AntCar project is to develop a new route-following method for mobile robots or intelligent vehicles in environments lacking satellite geolocation, such as urban canyons or indoor spaces. To achieve this, we used a robust, biologically constrained neural model inspired by ants, previously developed in simulation, to evaluate the familiarity index of a panorama.

This algorithm, used in visual compass, consists in determining the orientation of the maximum familiarity index with respect to the panoramas learned along a path. A car-like robot was fitted with a 220° panoramic camera. The algorithm corrects course using very low-resolution 44×44 pixel images (5°/pixel) inside and outside to determine the direction to follow along the previously visually learned path.

The results obtained are particularly reliable and reproducible. The biologically constrained neural model stores visual information sparingly and efficiently, so that visual memory has a very small memory footprint of just a few tens of kilobytes per kilometer traveled.