It’s well known (among the sort of people who think it’s a good idea to administer intelligence tests to ants) that navigating animals extract useful information from the shape of the skyline to guide their routes. In fact, if you replace the real skyline with a crude facsimile (scaled down but far closer to the ant), they are unfazed and carry on navigating to the same place they originally learned. We took advantage of this by transferring wood ant colonies to a lab and letting them learn to find food, guided only by a panorama made up of pieces of cloth pegged to the walls of a big white cylinder. We then tried to deduce exactly what the ants had learned by changing the shapes and looking at where the ants search for food: if we choose our shapes cleverly enough, the kinds of mistakes they make should let us know what rules they are trying to use.

 

One big question is whether ants segment a scene into component shapes, the way we do. When you look at the horizon, you don’t simply see it as a single undulating line; instead, without any conscious effort on your part, you see a mountain over here, a tall tree over tree, a stand of woodland in the distance (is that one object or many?). Do ants do something similar? Do they draw their mental dividing lines in the same places we do? And however they divide the scene into components, how do they store information about those shapes so that it can be used to guide them to their food?

 

We trained ants to find food at a point between several large skyline shapes, so that rather than heading toward a particular shape, they had to learn something about the relative positions of the shapes on either side of the goal. Then we removed or altered the shapes to see where they would go. Replacing the learned shapes with different ones in the same positions had no effect on the ants’ ability to find the food, so they were evidently able to apply their stored memories of the relative positions of shapes to a new panorama even though the shapes didn’t match at all.

 

When we trained them to head in a direction specified by the positions of a rectangle and a triangle and then tested them with just one of those shapes present, you might expect that they would just retrieve the memory associated with the one shape that remained; so if the food was usually found 30 degrees to the right of the rectangle and 60 degrees to the left of the triangle, and they were tested with just the rectangle, they might head 30 degrees to its right. In fact, this is not what happened, and the ants always seemed to take a compromise heading that didn’t match the feeder’s position with respect to any single component of the training panorama. The different tests we performed, and their analyses, get a little complicated, here, but the overall pattern of the results suggests that the ants have learned information about each shape (probably including the angles of the edges and some measure of its size or “visual mass”) along with its position relative to the feeder. When the ant encounters an ambiguous view, it can retrieve all of these memories each of which suggests a direction of travel. To anthropomorphise, the thought process might be: “if I’m looking at the rectangle I should turn 30 degrees to the right, but if I’m looking at the triangle, I should turn 60 degrees to the left”. These separate possible headings are then recombined according to some algorithm we don’t yet fully understand, to come up with a best guess compromise heading.

 

Not every memory is given equal weighting: it seems that how well the shape matches plays an important role (so if the test shape is a rectangle, the rectangle memory will play more important role in generating the compromise and the ant will head to the right, just not so far right as it normally heads, relative to the rectangle; but if it’s a triangle, the triangle plays a more important role and the ant heads further left, just not so far left as normal). The distance of each shape from the food also influences the eventual heading, with more weight given to shapes that are normally close to the food than those that are more distant. It’s likely that other factors also play a role and that the compromise heading is constantly updated to reflect the ant’s level of certainty about how its current view matches the stored memories.

 

Tom Collett suggested a useful analogy to visualize this procedure: imagine the ant wearing a loop of wire like a crown, with a few brightly coloured beads in certain locations around the ring. These beads match the memorized positions of the shapes on the skyline. To navigate toward the food, the ant turns until the beads all align with the shapes on the skyline and walks in that direction. When the skyline is changed, there is no way to perfectly align the beads, so the ant finds a position which in some way provides the least-bad fit. Perhaps imagine that the ant turns repeatedly this way and that, matching one bead to the shape on the horizon, realizing that the other beads are unaligned and turning again to match one of them, instead. It lingers the longest or returns most frequently to the positions in which the match is best, and quickly turns away when the match is bad, and in this way a sort of average direction of travel gradually emerges from the disorder. Ants do walk with a sinuous zig-zag path, so it’s not impossible that something like this repeated turning and matching really does take place, but we suggest it only as an analogy to show how different, incompatible memories can be reconciled in the brain to come up with a best guess direction of travel.

 

This process, in which the brain simultaneously comes up with several independently calculated estimates of where to head, and only combines them to produce a single direction of travel at a much later stage of processing, seems to be a general principle of insect navigation. For example, in addition to their visual memories, ants use a process called “path integration” to keep track of their position and can use it to plan a direct route home, even if their outbound route was very tangled. Many species of ant can also follow scent trails laid by their nestmates. These different ways to plan a route are kept separate in the brain and play a larger or smaller role in the eventual choice of travel direction depending on the context and the ant’s estimates of their reliability: when the path integrator gives a clear indication of the homeward direction, ants will generally follow it even if the skyline indicates they are on the wrong track, but when they are close to home it is less reliable and visual cues play a bigger role; meanwhile, an inexperienced ant will often follow pheromone trails even if they conflict with the other sources of information, but experienced ants rely on their own private knowledge (like visual memories or path integration) in preference to following trails.

 

*These authors contributed equally to the study.