Route Learning in Ants and Robots

In a close cooperation with the Neurobehavioural group at the Institute of Zoology the visual mechanisms underlying polarized skylight navigation, path integration (dead-reckoning) and landmark navigation (visual piloting) in the Saharan desert ant Cataglyphis are studied.

Due to the flat environment these ants live in it is impossible for them to navigate only map based by visual landmarks. They do not use chemical path signalling mechanisms as well, probably because they are scavengers of dead insect corpses which are evenly distributed in the area and thus there is naturally no particular food-source. So they must combine several amazing mechanisms in order to navigate accurately in the desert.

In path integration, the insects determine somehow their distance and direction from the nest incrementally over the entire journey. Therefore, we term this computed vector pointing at home a global vector. Due to errors it has to be aligned by a compass, in this case based on the pattern of polarized skylight.

The landmark panoramas are employed particularly for generation of snapshots near the nest. The snapshot-matching model assumes that the views on which the matching process operates are horizontal images comprised of dark and bright sectors corresponding to the landmarks and the inter-landmark gaps, respectively. Former experiments performed on a robot (Sahabot 2) showed that the snapshot-matching algorithm, in conjunction with the polarized skylight compass used for aligning the views, worked and produced reasonably accurate homing trajectories. See the Saharan Ants and Robots Project which is in cooperation with this project.

Experiments have shown that snapshots are generated also on the road when landmarks are visible. In most cases, however, the global vector is dominant in guiding the animal.

There is a particular experimental setup developed by Sonia Bisch-Knaden in which the ants evidently are not able at first to determine the global vector back home properly. This setup consists of a V-shaped barrier which the animals can cross the way from the nest to the artificial food-source, but not the way back. She observed that they need several trials to learn to adjust their homepath. Each ant learns a particular path and does not change it further after a certain number of trials. For this adjustment they apply a so called local vector attached to a certain location on the road, in this particular case to the end of the barrier. The question was to determine by which means this local vector is stored. Is it stored egocentrically, as the angle to the direction they run just before, i.e. the direction of the barrier? Or is it stored geocentrically, as the absolute direction, i.e. the angle relative to the skylight compass? Could it be finally a combination of both, which is suggested by experimets with channels?

Bisch-Knaden has found only egocentric local vectors when the ants run along the barrier for several meters (showed at the ZNZ Symposium 2000: Poster No 138).

The only setup which showed a local geocentric vector were experiments with channels, which showed preferentially egocentric vectors after quitting long channels, and more geocentric vectors after short channels (Collett et al. 1998).

I used the described barrier setup to quantify the learning phase at the beginning and to search for an eventual geocentric local vector recalled at the end of the barrier after running only a short path alongside.

My results are preliminary to this date, but they seem to show that there actually is only a egocentric local vector, i.e. it is only the angle to the barrier.

From the methodological point of view I felt that is was time to change the traditional system of drawing the ants trajectories on paper for scanning and digitizing them later in the lab as it was used for the last 30 years. I thought that technology now is at the point that it should be possible to spare the paper-step and digitize the data right in the field.

So in summer 2000 I used a device consisting of a notebook and a graphics tablet, powered by a solar panel I was wearing on my back to track the ants trajectories. The temperatures as high as 45 °C in the shade and huge amounts of salt and dust in the air, together with the fact that is was impossible to prevent the notebook from direct sunshine raised the question if it would not crash after a time. But the system was doing well for three month, despite the harsh environmental conditions, as you can see in the pictures below.

Next year I will focus on navigation in complex arrays of landmarks and try to investigate the interactions of the different navigation mechanisms on each other and with other systems like learning and memory, attention and decision processes. Therefore, in a broader sense, contextual information processing will be a major issue in this work.

The generated hypotheses will then be tested in simulations and autonomous agents, particularly using Sahabot 2.

In the long term we expect as a result a better understanding of navigation and contextual information processing in general. Another result will be a better understanding of the synthetic methodology in interdisciplinary projects.