Learning walks and landmark guidance in wood ants (Formica rufa)

Many animals recognise significant sites in their environment in terms of particular sets of visual features and rely on their memory of these features when returning to a familiar place (for reviews, see Gallistel, 1990; Healy, 1998). Such navigational information needs to be acquired rapidly, and bees and wasps have long been known to perform specialised learning flights on first leaving their nest or a newly discovered feeding site (for reviews, see Zeil et al., 1996; Collett and Zeil, 1996). These insects learn the layout of landmarks in the immediate vicinity of a site sufficiently well during a single flight that they are able to return there (e.g. Tinbergen, 1932). We describe here the properties of a behaviour pattern in walking wood ants that may play a similar role in learning.

When wood ants leave a newly found drop of sucrose located at the base of a conspicuous landmark, they periodically turn back, fixate the landmark, and walk a little way towards it. Close to the landmark, where the landmark’s image on the ant’s retina transforms rapidly, inspections occur often. Their frequency decreases as the ant moves away and the landmark’s image transforms more gradually with distance. One possible reason for the decrease in the number of inspections with increasing distance is to allow the ant to sample views of the object at equiangular intervals (Judd and Collett, 1998). If an ant were to operate such a sampling strategy with an uncertain separation between sucrose and landmarks, it would need to control the distribution of its inspections visually. The distance between one inspection and the previous one could, for instance, be set by the image transformation that an ant observes during its preceding inspection. Another explanation for the distance-dependent decline is that the ant requires more views when it is near to its goal than it does when further away, and it simply reduces the probability of inspections as it travels away from its goal, irrespective of its distance from a landmark.

To distinguish between these alternatives, we have varied the separation between feeder and landmark. This manipulation also allows us to ask more precisely where ants look during inspections. Do they primarily fixate landmarks or the feeding site?

The results of these experiments combined with earlier work on landmark guidance in desert ants (Wehner et al., 1996) provide the basis for a model of image matching that emphasises the ants’ predilection for keeping discrete landmarks in their frontal visual field.

Colonies of queen-right wood ants (Formica rufa L.) were collected from local woodland and housed in large tanks in the laboratory at 20–25 °C. Ants were fed frozen crickets weekly and were allowed to forage for sucrose, either in a rectangular arena (96 cm ×72 cm), which they reached by means of string and plastic tubing running from the holding tank to the arena, or on a shelf (91 cm ×70 cm) above the nest, which could be reached by a removable string drawbridge. The floor of the arena and shelf were surfaced with white laminated plastic that had been roughened slightly with sandpaper to prevent the ants slipping and against which they were clearly visible. The plastic was cleaned regularly with alcohol to remove pheromone trails. The arena and sometimes the shelf were surrounded by curtains. In most experiments, the holding tank with its shelf was kept in one corner of a windowless laboratory with a minimum of distracting features within the ant’s field of view.

Experiments were performed while ants foraged singly for sucrose. An individual ant, marked with paint when necessary, was placed on to the string leading to the foraging area. A few drops of sucrose were placed either directly onto the plastic surface or onto a piece of microscope slide. The food was approximately 70 cm from the entrance to the shelf or the arena. According to experimental requirements, one, two or no landmarks were positioned at a given distance from the sucrose. The return walks from the feeder to the nest, and sometimes approaches to the feeder, were recorded at 50 frames s⁻¹, using a video camera fixed above the arena. The need to resolve the ant’s longitudinal axis meant that the area viewed by the camera was limited to approximately 50 cm ×40 cm.

Video tapes of the ants’ paths were analysed using a commercially available software package which had been modified to provide the horizontal orientation of the ant’s body axis as well as its position on each frame (Judd, 1998).

One similarity between the learning flights of bees and wasps and the ‘learning walks’ of wood ants is that both are typically performed when a site is new to an insect or has not been visited recently (e.g. Zeil, 1993; Lehrer, 1993). On an ant’s first return from a feeder, there may be many backward turns, defined here as path segments of at least 0.5 cm in which the ant faces the landmark directly. As the ant becomes more familiar with the feeder and the path leading to it, backward turns become less frequent. In general, inspections of the landmark are rarely found after approximately three trips to the same site. In Fig. 1, we show the number of backward turns of one ant on its first 40 departures from a cone contiguous to a sucrose feeder. Backward turns per visit decrease to zero over the first day on which the ant made 16 trips. They recur at the start of day 2 (trip 17) and of day 3 (trip 31). The ant was then prevented from foraging on the shelf for 20 days, and the number of backward turns increased dramatically on the first two visits after the break.

Single ants approached a feeder marked by a black cylinder that was positioned either 20 cm beyond the sucrose or immediately behind it. We recorded the ant’s first two or three return trips to the nest after it had discovered the feeder. Sometimes the ant left the drop of sucrose and then returned all the way to the drop. These movements were considered to be part of feeding and were excluded from the analysis. The returns to the nest were screened for segments in which the ant moved towards the feeder rather than the nest and looked in the approximate direction of the landmark or feeder. This was achieved by identifying segments in which the ant was both oriented within ±90 ° of the direct vector from the drawbridge to the feeder and in which the direction of the centre of the landmark or feeder from the ant was within ±45 ° of its longitudinal axis. Histograms were plotted of the distance of the ant from the feeder for each frame of these recorded turn-backs pooled over all returns.

As reported previously (Judd and Collett, 1998), the time spent looking backwards decreases with increasing distance from the feeder. However, the exponential fall-off is much the same whether the cylinder is 20 cm (Fig. 2A) or immediately (Fig. 2B) behind the feeder. If the shape of the distributions were the outcome of a strategy of sampling the cylinders at equiangular intervals, inspection time should decrease more steeply when the cylinder is adjacent to the food than when it is 20 cm behind it. To compare these conditions statistically, we measured the distance of the beginning of each inspection from the feeder. A Mann–Whitney U-test was applied to determine whether there was a significant difference between the distances of the inspections with the cylinder next to the feeder (23.79±9.76 cm, mean ± S.D., N=85) and the inspections with the cylinder 20 cm from the feeder (22.99±9.97 cm, mean ± S.D., N=80). The standard normal deviate was 0.665, with P<0.253, one-tail (U=3196). The decrement seems to be governed more by the ant’s distance from the feeder than by its distance from nearby landmarks.

Where do ants look during inspections if the landmarks and feeder are separated? We investigated whether both landmarks and feeder are fixated, whether ants look preferentially at objects close to the feeding site that would provide accurate guidance cues and, finally, how ants move while fixating landmarks.

Two examples of an ant’s return to the nest after feeding are shown (Fig. 3) for an arrangement of two cylinders placed either side of and behind a feeder. The thickened portion of the trace shows when the ant ‘fixated’ within ±20 ° of the centre of the landmarks or the feeder. In both examples, the ant clearly fixates the cylinders more than the feeder, although in some positions it is unclear whether the ant is looking at the feeder or a cylinder.

To determine inspection preferences more quantitatively, we analysed the positions of the landmarks and feeder on the ant’s retina during segments of homeward trips in which the ant was moving towards the feeder rather than the nest. The assumed horizontal position of a feature on the ant’s retina was taken as the viewing direction of the centre of the feature relative to the ant’s longitudinal axis. The distributions in Fig. 4A illustrate the ‘retinal’ position of landmarks while the ant was oriented within ±90 ° of the direct vector from the arena entry to the feeder.

The landmarks were two black cylinders located at different distances from the feeder and scaled so that they had the same image size when viewed from the feeder. The bimodal peaks in the two distributions of Fig. 4A imply that ants looked at both cylinders. The ants showed no obvious preference for the cylinder closer to the feeding site, and there is little indication of a third peak corresponding to fixation of the feeder. Fixation seems to be primarily of conspicuous visual objects with little attention being paid to the feeding site. The lack of fixation of the feeder is not just because local conspicuous objects distract the ant’s gaze from the feeding site. It persists when ants forage at a feeder with no objects on the shelf and a curtain enclosing three sides of the shelf to prevent the ant from seeing beyond the tank. In this case, the retinal distribution of feeder positions is very broad (Fig. 4B).

During inspections, ants tend to approach a cylinder while fixating it. The path segments travelled during fixation of a cylinder are shown in Fig. 5. The data are taken from several different conditions, and the feeder positions are therefore not shown. As expected, the superimposed paths generate a radial pattern centred on the cylinder.

A similar mode of inspection is found on approach trips to the feeder. The sample trajectory in Fig. 6 is of an ant that has foraged for several days at a feeder positioned between an inverted cone and a narrower cylinder. During this approach, the ant first fixates the more conspicuous cone and then the cylinder. The histograms in Fig. 6A–C show the retinal positions of the cone (Fig. 6A) and the cylinder (Fig. 6B,C) during 15 approaches. When the ant is further than 20 cm from both the cone and the feeder, it tends to fixate the cone (Fig. 6A), so that the cylinder is viewed off to the right (Fig. 6B). When the ant is closer, it fixates the cylinder (Fig. 6C) rather than the cone.

There are strong indications, but no conclusive evidence, that the acquisition of landmark views occurs on ‘learning’ walks such as those described here. First, inspections of landmarks are most frequent when learning is expected to take place, that is after food has recently been discovered or when there has been a long time interval between foraging trips. Second, the ant, during its inspections of an object, obtains retinal views of the object that correspond to those that it tries to match on later returns (Judd and Collett, 1998).

Interesting differences are seen between learning flights in wasps and bees and the putative learning walks of ants. A major feature of learning flights is that the insect moves along arcs centred roughly on the goal (Zeil, 1993), a movement pattern that seems to be designed to provide parallax information about the distance between the goal and landmarks. Second, bees and wasps fixate the goal when changing direction at the end of arcs and when their rotational velocity is low (Collett and Lehrer, 1993). They may acquire views for image matching at these points.

Ants, in contrast, primarily fixate landmarks and move towards them, with no obvious arcs around the feeder. The small pool of sucrose at which the ant feeds does not appear to be a prime object of fixation. To fixate such an inconspicuous target, the ant would need to use its path integration system to specify the feeder’s direction. It could be argued that path integration becomes more inaccurate indoors, where skylight cues giving direction (Wehner and Rossel, 1985) are not available, and that under these conditions it is difficult to fixate the feeder. There are reasons for doubting such an explanation for ignoring the feeder. Some ants can obtain compass direction from canopy cues (Hölldobler, 1980), and there is evidence that the formicine Tetramorium caespitum navigates indoors by path integration as accurately as it does when provided with a skylight compass outside (Shen et al., 1998).

During normal foraging, Formica rufa follow fixed routes guided by visual landmarks (Rosengren, 1971; Rosengren and Pamilo, 1978). On the assumption that learning does occur during inspections on return trips, what might the details of learning walks tell us about landmark guidance, noting, of course, that the present experiments were restricted to a very small area compared with the ant’s usual foraging range? First, the ants’ frontal fixation of landmarks suggests that they learn about discrete objects using their frontal visual field and that such views help guide them towards the food. This agrees well with the frontal fixation of landmarks that we observe on regular foraging trips. Second, a concentration of landmark fixations close to the goal suggests that ants may normally pin-point the goal using frontal views of landmarks.

A reliance on landmark-centred views in image matching might explain an interesting difference that has been found between the search patterns of honeybees and desert ants (Cataglyphis fortis) when they are tested with arrays of landmarks that have been transformed from their normal arrangement. Bees were trained to a feeder marked by three cylinders and then tested with an expanded array in which each cylinder was placed at double the training distance from the feeder. Bees, in such tests, searched where the viewing directions of the landmarks were the same as usual, even though the apparent size of each landmark at the search site was then half the training dimensions (Cartwright and Collett, 1983). Ants trained with a triangular array of landmarks centred on the nest were challenged with a similar expanded array. In this case, the search area was only circumscribed if each individual landmark was proportionally enlarged, enabling ants to find a position where the two-dimensional view matched in all respects that experienced at the nest during training. Expansion of the arrangement without enlarging the landmarks generated a diffuse search area (Wehner and Räber, 1979; Wehner et al., 1996).

One possible explanation of this difference is that ants place a strong weighting on information in their frontal visual field, and thus require a correct frontal view of landmarks, which they fixate. A broad search area with expanded arrays but without enlarged landmarks would arise if ants tried to achieve correct frontal views of all landmarks. In this scenario, an ant would approach one landmark, turn to check the appearance of another and walk towards it in a search that continues without a stable endpoint. It might be worth examining the search paths of Cataglyphis fortis with this suggestion in mind.

We tested the plausibility of this idea using a very simple computer simulation. Simulated ants before leaving a goal to which they wished to return faced each landmark in turn and, while facing it, took a ‘snapshot’. Each snapshot consisted of a 360 ° narrow horizontal strip on which was imprinted a view of the landmarks. When started at a position away from the goal, the simulated ant selected and faced one landmark. It then calculated a matching score between its current view and each of the snapshots. The score was obtained by summing the angular difference between the centre of each landmark in the retinal image and the centre of the nearest landmark in the snapshot. The snapshot attracting the smallest score was engaged. In this part of the process, we have arbitrarily chosen not to weight the frontal retina. In the next stage, the landmark fixated in the frontal retina dominated the simulated ant’s behaviour: the ant moved towards the fixated landmark until its image was larger than its partner in the snapshot. The simulated ant then stopped and repeated the cycle. It turned to fixate a landmark at random, selected a snapshot as before, and moved to improve the match.

This algorithm guides simulated ants to a goal at the centre of a triangular array, as in the experiments reported by Wehner et al. (1996) (Fig. 7A,B). It also replicates their finding that expansion of the array coupled with enlargement of individual landmarks leads to a circumscribed, but broadened, distribution of endpoints (Fig. 7D). Again, like those of real ants, the simulated paths do not settle on an endpoint if the array is expanded but the landmarks are not enlarged (Fig. 7E).

We tried several more-elaborate variants of this scheme and all gave qualitatively similar results. The algorithm is not meant to be a complete simulation of landmark guidance at the nest. It does not attempt to link snapshots to compass direction, and it often fails to define a unique goal when individual landmarks are not distinguishable. For instance, if the goal is placed asymmetrically within an equilateral triangular array of three identical cylinders, the simulation generates three potential goal sites, and the site to which the ant is drawn will depend upon its starting point (Fig. 7C).

Why might ants have different strategies from wasps and bees? A possible reason lies in the ability of wasps and bees to fly obliquely or sideways, as well as forwards – a competence that is not shared by walking ants. This flexibility allows the insect to change flight direction without rotating. It can thus adopt a constant orientation in the last phase of its approach to a goal and adjust its path exclusively through translational manoeuvres (Collett and Rees, 1997). Image matching is easier to implement if changes in direction can be accomplished without rotation (Cartwright and Collett, 1983) because the dimensionality of the search space is reduced. With no ability to walk sideways, ants need an alternative movement strategy to limit their search when image matching. One possibility is to aim successively at different landmark features close to the goal. If ants store images when fixating landmarks frontally, they then ‘know’ that images must be matched in the same manner.

We thank Paul Graham for comments on an earlier draft. Financial support came from the BBSRC and the Human Frontier Science Program.

We hear from Ralf Möller that he and his colleagues have developed a somewhat different algorithm to mimic the search patterns of desert ants in expanded arrays. (Möller, R., Lambrinos, D., Pfeifer, R. and Wehner, R. (1999). Do desert ants use partial image matching for landmark navigation? Proc. Neurobiol. Conf. Göttingen, in press.)