A desert ant's memory of recent visual experience and the control of route guidance

Abstract

Insects such as desert ants learn stereotyped visual routes between their nests and reliable food sites. Studies here reveal an important control element for ensuring that the route memories are used appropriately. They find that visual route memories can be disengaged, so that they do not provide guidance, even when all appropriate visual cues are present and when there are no competing guidance cues. Ants were trained along a simple route dominated by a single isolated landmark. If returning ants were caught just before entering the nest and replaced at the feeder, then they often interrupted the recapitulation of their homeward route with a period of apparent confusion during which the route memories were ignored. A series of experiments showed that this confusion occurred in response to the repetition of the route, and that the ants must therefore maintain some kind of a memory of their visual experience on the current trip home. A conceptual model of route guidance is offered to explain the results here. It proposes how the memory might act and suggests a general role for disengagement in regulating route guidance.

1. Introduction

Insects such as the desert ant Cataglyphis fortis use a variety of memory-based guidance systems to learn and follow habitual routes between their nest and a food source. When an individual encounters a familiar view, the appropriate visual memories that have been acquired on previous trips are generally activated to provide guidance on the current trip. How individuals ensure that they use the appropriate memories at the appropriate times is still poorly understood. Part of the solution is likely to be that insects associate memories with specific goals. A nest-bound individual carrying food, for instance, will use only nest–ward memories [1–3]. Several studies have also suggested a priming effect, acting at a finer scale, where the visual scene at one location may bias the use of memories at an immediately subsequent location [4,5]. This sequential priming effect appears to be short-lived, disappearing within at most 5–10 s [6,7]. This study reveals an additional regulatory process involved in the use of visual route memories.

If an individual travels a path repeatedly, then it acquires visual-based route memories of its habitual heading directions along the path [8–12]. By rotating so as to align the current retinal view with retinotopic visual memories acquired on previous trips, an individual will automatically travel in the habitual directions from points along a route [8–15]. This widely used mechanism can be termed ‘alignment image-matching’ [11,16]. It allows individuals to follow stereotyped and often idiosyncratic routes [17–19]. In addition to view-based guidance, C. fortis uses guidance from path integration (PI) [20]. Whenever an individual is out foraging, it monitors the directions and distances that it travels, updating a memory of the cumulative distance and direction travelled from the nest—its ‘PI state’. To use PI information for guidance, an ant effectively subtracts its current PI state from the remembered PI state of its goal. The resultant of this subtraction—the ‘PI output vector’—encodes the direction and distance that must be travelled to reach the goal [21].

When returning to its nest from a familiar food site, an ant will be guided home by both route memories and PI [14]. If the ant is captured just before entering its nest, then its PI state will continue to indicate that it has reached the nest, even if the ant is replaced back at the familiar food site (reviewed in [22]). This means that once the ant has been displaced, the PI output vector will no longer point towards the nest: on release, the ant is said to have a ‘zero vector’ [20]. Nonetheless, as long as it retains its homeward motivation (which it generally does so long as it has food to take home), a zero-vector ant will use its visual memories to recapitulate the route it just took [11,17–19]. In the current study, ants were trained along a route that passed a single dominating landmark in an otherwise largely bare environment (figure 1a and the electronic supplementary material, figure S1). From the feeder, the ants followed homeward paths that curved around either side of the cylinder (figure 1b). Unexpectedly, if ants were caught just before arriving at the nest and then displaced back to the feeder to repeat this route, they often interrupted their recapitulations on the approach to the landmark (figure 1c). They then appear confused, exhibiting a period of search-like behaviour before resuming their routes home. The ants stop using their route memories even if the appropriate visual cues are present and if there is no competition with other forms of guidance.

Figure 1.

Intervals of confusion during recapitulation trajectories. (a) Photo of experimental area viewed along the homeward route from the feeder. The feeder is marked by the bottle, and the nest is 7 m behind the black cylinder landmark. (b) Initial homeward trajectories from the feeder. The circle represents the location of the large cylinder, the triangle the position of the feeder and the cross the position of the nest. (c) Recapitulation trajectories from feeder (n = 18). The dark segments show the intervals, judged by eye, between the start and endpoints of confusion. (d) Start and endpoints of confusion during recapitulation trajectories. Recapitulations released from (i) y = −4 m (trajectories shown in figure 2c), (ii) y = 0 m (trajectories shown in figure 1c), (iii) y = +2 m (trajectories shown in figure 2a). Solid line on the left of each route schematic extends from release point (indicated by arrowheads) to the mean start point. A faint dotted line extends to the maximum start-point. Solid line on the right of each route schematic extends from the mean endpoint of the confusion to the nest. The faint dotted line extends from the minimum endpoint to the mean. Minimum and maximum start and endpoints are indicated by short horizontal lines. (Online version in colour.)

An anthropomorphic explanation for the confusion might be along the lines that an individual is surprised to find itself again approaching a landmark that it has already passed, or perhaps that the repetition violates the individual's expectation of what it should see or do next. In a human, these explanations would probably include a role for episodic memory, with an individual remembering that he or she has just been at a particular place, and so should not be there again. An ant-centric explanation should probably not invoke something as sophisticated as episodic memory. However, the subsequent manipulations show that it does involve some kind of memory (lasting longer than the few seconds of the short-term memory involved in sequential priming) of the ant's visual experience or actions along the way home. This memory appears to form part of the regulatory processes that control when the route memories are used.

2. Methods

Experiments were carried out with a single C. fortis nest at a particularly flat and featureless location on the island of Djerba, Tunisia during June 2006 and June 2008. In both years, ants were trained to a feeder (a slice of watermelon) 14 m north from their nest. A black cylinder 70 cm high and 40 cm wide was placed halfway between the nest and feeder. In 2006, the cylinder was on the direct line between the nest and feeder, whereas in 2008, it was placed 1.5 m to the side of the direct line. In 2006, but not in 2008, the feeder was marked by a plastic water bottle filled with sand. Fine string was laid out in a grid of 1 m squares across the area so that the trajectories of ants could be transcribed by hand. For the sake of simplicity, the feeder–nest axis is referred to as the y-axis (figure 1b), with the feeder at (0,0) and the nest at (0,14).

Approximately 80–120 ants were at the feeder at any time, suggesting that the total number of foragers visiting the feeder was over 200. An ant to be tested was selected at random from those preparing to leave the melon. The selected ant was given a crumb of biscuit to ensure a homeward motivation, and to allow easy identification for capture near the nest. After an ant had been tested, it was kept in a vial until the end of the testing session. Thus, no ant was tested twice in the same morning or afternoon. Because of the large numbers visiting the feeder, few, if any, ants would have been tested more than once in any conditions.

A series of manipulations explore the roles of PI, and of the repetition of visual stimuli, in eliciting the apparent confusion during the recapitulations. In ‘recapitulation tests’, an ‘initial’ homeward trajectory from the feeder was recorded, and then the ant was captured just before it arrived at the nest. It was then displaced to one of three locations from where its ‘recapitulation’ trajectory was recorded. These recapitulation tests were interleaved during days 12–15 after training started. In 2006, a series of ‘extension tests’ were interleaved during days 16–17. Ants were displaced from the feeder to one of two locations to run a single long trajectory back to the nest. Finally, pairs of ‘combined tests’ were interleaved during days 17–18. In these, ants travelled 7 m, either on or off of the route, before being displaced to (0,−11) from where they travelled the 25 m back to the nest. In 2008, the data were collected after 3–11 days with the landmark and feeder in place. Towards the end of this period, the levels of confusion appeared to decrease (electronic supplementary material, figure S6). Because, in 2008, the confusion levels changed with time, the 2008 analyses are limited to the subset of the collected data for which comparisons can be made within the same testing sessions.

The trajectories were digitized using GraphClick (Arizona Systems) and analysed using Matlab. The intervals of confusion shown in the figures were the segments, judged by eye, that did not appear to be part of a smooth arc towards the nest. The statistical analyses used instead an objective measure of the path length between the feeder and the landmark: the ‘excess path length’. The excess path length between two points was calculated as the total path length between the points, minus the distance between the points. If an ant did not resume its path back to the nest, then the excess path length was assigned to be 20 m (slightly larger than the largest measured excess path length). For comparisons involving the recapitulation trajectories starting at y = +2 m, the start point was the first time the ant crossed y = 2 m. For the remaining comparisons, it was the first time it crossed y = 0 m. The endpoint was the last time y = 7 m. All data are given as mean ± standard deviation. Because the data are not normally distributed, all statistical comparisons used the non-parametric Mann–Whitney–Wilcoxon rank sum test (Matlab's ranksum test). Correlations between trajectories were calculated from the x-coordinates when y = 2 (before the confusion) or y = 11 (after the confusion), and used a linear least-squares fit.

3. Results

(a). Switching route memories off and on

When homing ants were caught just before entering the nest and replaced at the feeder, their trajectories (figure 1c) started out as recapitulations of their initial trajectories (figure 1b). After travelling a path length of 4.1 ± 1.0 m, however, 14/18 ants abruptly interrupted the recapitulation of their route. These 14 ants then appeared to search for 7.8 ± 5.4 m before resuming their paths home. Both before and after the period of confusion, there is a strong correspondence between an ant's initial and recapitulation trajectories: before the period of confusion, 10/14 ants direct their trajectories to the same side of the landmark as on the initial trajectory, and after the period of confusion, 11/14 ants do, with r² = 0.58 and r² = 0.40, respectively (correlations shown in the electronic supplementary material, figure S3).

As in other studies [17–19], the displaced ants started out by recapitulating these routes. Why then do they break off and temporarily stop using their route memories? The interval of apparent confusion does not occur as a result of competition with a stronger guidance cue, as the route following gives way to search and not to some alternate form of guidance. Does the interruption occur instead because there is a section of route where visual cues are too ambiguous or too weak to provide guidance? One piece of evidence against such a gap in coverage is that, on average, the point at which ants resume their routes (y = 3.5 ± 2.2 m) is no further along the route than the point where they broke off (y = 3.5 ± 1 m), with half (7/14 ants) resuming their routes at an earlier point. This absence of a consistent direction of movement during the confusion suggests too that the behaviour really is some kind of search [23].

Further evidence against a gap in the route memories is provided by a pair of additional manipulations that shifted the release sites for the recapitulation trajectories (figure 1d). From a release site at y = +2 m, only 11/20 ants show obvious signs of confusion (figure 2a). The confusion in these 11 ants starts after ants have travelled 3.1 ± 1.7 m, and ends after 2.9 ± 2.9 m of search. Again, the locations at which the confusion starts and ends are approximately equal: it starts at y = 4.6 ± 1.5 m and ends at y = 4.5 ± 2.0 m (see also the electronic supplementary material, figure S2). If the zero-vector ants are instead released from y = −4 m, then they again first head towards the nest. From the release point, the retinal size of the landmark is only 64% of its size at the feeder, but the view is evidently otherwise similar enough to provide guidance. 22/24 of these recapitulation trajectories show an interval of confusion that starts after ants have travelled 4.7 ± 3.0 m (figure 2c). The path length travelled before the confusion is not significantly different from that in the recapitulations from the feeder (p = 1, Z = 0). In 16/22 of these ants, the confusion ends after 11.9 ± 7.6 m of search (in some cases, including an intermediate stretch of directed travel)—the remaining six do not recommence the route. The locations of these start and endpoints of confusion are at y = −0.1 ± 2.4 m and y = 0.8 ± 2.8 m. The locations of these start points are significantly different from those for the recapitulation trajectories from both the feeder (p < 0.01, Z = 4.0) and the +2 m release site (p < 0.01, Z = 3.9), and the endpoints are significantly different from the +2 m release site (p < 0.01, Z = 2.7), but not the feeder (p = 0.06, Z = 1.8). Thus, although, in all conditions, the break in route following occurs during the approach to the large cylinder, it is not yoked to specific retinotopic views. As in previous studies [11,17–19], the ants must therefore have visual memories of the entire route that they can use even without support from PI.

Figure 2.

The effect of repetition. Left-hand panels show conditions in which ants are recapitulating the route. Right-hand panels show conditions, matched for the ant's PI state, in which the ants travel along the route only once. The arrows indicate the point at which the PI state is zero. The dark segments show the intervals of confusion. The boxed insets illustrate the manipulations schematically: dotted lines represent the paths travelled from first release to final capture; triangles, circles and crosses represent the feeder, landmark and nest. (a) Recapitulation trajectories from 2 m in front of feeder (n = 20). (b) Extended homeward trajectories for which ants are displaced from the feeder to 12 m behind the feeder (n = 8). The 14 m segment from −18 until –4 m is not shown. (c) Recapitulation trajectories from 4 m behind the feeder (n = 24). (d) Extended homeward trajectories for which ants displaced from the feeder to 18 m behind the feeder (n = 19). The 8 m segment from −12 m to −4 m is not shown. (e,f) Trajectories matched for PI output vector on release. (e) The ants are first allowed to travel the first 7 m along the route, before then being caught and released at −11 m to return all the way home (n = 18). (f) The ants are first released to travel 7 m from −11 to −4 m (i.e. not along the route) before being caught and re-released from −11 m (n = 14).

At all of these release sites, the ants express their route memories immediately upon release. After travelling approximately 3–5 m towards the landmark, there is then often a sudden switch in their behaviour as they appear to stop using their route memories. After a brief period of search, almost all ants then show another switch in their behaviour, resuming their homeward routes. These switches during the recapitulation trajectories do not occur because of the disappearance or reappearance of the familiar visual cues. They appear to reflect some sort of changing navigational state in the ant.

(b). A memory of visual experience on the current trip

What produces the differences between the initial trajectories, which are always uninterrupted, and the recapitulation trajectories, which often show an interval of confusion? Because there are no changes to the environmental stimuli, it must be something internal. The most obvious explanations would be related to PI. During the initial trajectories, both route memories and PI provide guidance in the homeward direction. On the recapitulation trajectory, however, an ant's PI system no longer provides guidance in the feeder–nest direction [22]. If the confusion is simply due to the absence of guidance from PI, then manipulating an ant's PI state before it started its homeward route should produce the same level of confusion as making the ant repeat the route. A further set of manipulations showed that this is not the case.

Ants were captured from the feeder and then displaced so that they ran ‘extended’ homeward trajectories from either 12 m or 18 m behind the feeder. From these two release sites, the PI states would reach ‘zero’, indicating the position of the nest, after the ants had travelled 14 m in the nest-wards direction. This zero-point occurred either at y = +2 m or y = −4 m—i.e. at the same points where the zero-vector ants were released for the recapitulation tests (figure 2a,c). Of the ants released at y = −12 m, just 1/8 ants shows any obvious signs of confusion (figure 2b). In contrast, when ants are released from y = −18 m, so that the PI state reaches zero 4 m before reaching the feeder, some confusion is evident in 13/19 ants (figure 2d). The amounts of confusion can be compared quantitatively between conditions using a measure of the excess path length on the approach to the landmark (see §2 Methods for details. Distributions are shown in the electronic supplementary material, figure S4). Using this measure, comparisons between the two sets of extended trajectories suggest that the PI state may have some effect on the amount of confusion (n.s. p = 0.075, Z = 1.8), but it does not fully explain the amounts of confusion in the recapitulation trajectories. Levels of confusion are greater in the recapitulation trajectories than in the PI-matched extended trajectories (n.s. p = 0.18, Z = 1.3 for y = +2 m conditions; p < 0.001, Z = 4.0 for y = −4 m conditions).

Equally high levels of confusion in recapitulation trajectories are observed even if ants initially travel only the first half of the route. In a pair of further manipulations, ants were allowed to travel 7 m in the feeder–nest direction before being captured and displaced to y = –11 m from where they travelled back to the nest. Again, the manipulations ensured that the PI zero-point occurred at y = −4 m. In one group, the first 7 m was from the feeder—i.e. along the route (figure 2e). The degree of confusion in this group was not significantly different from the y = −4 m recapitulation test (p = 0.23, Z = 1.2). In the other group, the first 7 m was not along the route—ants were displaced from the feeder so that they travelled from y = –11 to –4 m before being recaptured (figure 2f). The confusion in this group was significantly less than in the group that had travelled the first half of the route (p < 0.01, Z = 3.0; distributions shown in the electronic supplementary material, figure S5), but was no different from in the extended trajectories from y = –18 m (p = 0.1, Z = 1.6). The high levels of confusion must therefore be caused by the repetition of the approach to the landmark. Furthermore, the effect of repetition means that the ants must have maintained, through the capture and displacement procedure, some kind of memory of having already travelled towards the landmark.

(c). The repetition of route memories

What aspect of the recent visual experience is encoded in the memory? A first stab at this question was made by examining whether the confusion is evoked specifically by an ant's repetition of its route, or by something more general related simply to approaching the landmark. In a slightly modified route, the cylinder was placed asymmetrically 1.5 m to the left of the direct homeward path, so that the habitual routes would all pass on the same side. As in the earlier study, if initial trajectories were along the habitual route, then the recapitulation trajectories again showed high levels of confusion—here after 2.2 ± 1.5 m (figure 3a).

Figure 3.

The specificity of experience along the route. (a) Recapitulation trajectories during days 3–5 (n = 14). The grey polygon shows the envelope encompassing the initial trajectories. (b) Initial homeward trajectories from four novel release sites (I–IV, indicated by circles) and from the feeder. (c) Histograms of excess path lengths after initial trajectories on (i) the unaccustomed side (n = 15) and (ii) the accustomed side (n = 14) of the landmark. Trajectories are grouped into those from the release sites (II and III) near the bifurcation point (open) and from the release sites (I and the feeder) approximately 2 m away (solid). Because of low sample sizes, the statistics were performed on the pooled data. Trajectories are shown in the electronic supplementary material, figure S7a. (d) Histograms of excess path lengths after initial trajectories (i) off of the route, from release site IV and (ii) from the feeder. Trajectories are shown in the electronic supplementary material, figure S7b.

Initial trajectories by ants released from (−2,0) and further to the right generally passed on the accustomed side, whereas those from (−2.5,0) and further to the left tended to pass on the unaccustomed side of the landmark (figure 3b). On the unaccustomed side, although the appearance of the landmark itself was the same, it was seen in the opposite side of the visual field and against the opposite half of the distant panorama. Two pairs of initial release sites were used to compare the effects of travelling on the accustomed versus the unaccustomed side of the landmark. At the end of these initial trajectories, the ants were recaptured and displaced back to the feeder. Zero-vector ants in all groups then showed some confusion (electronic supplementary material, figure S7), but the excess path lengths were significantly higher if the initial trajectories passed on the accustomed side of the landmark (p < 0.05, Z = 2.2; figure 3c). The high levels of confusion thus appear to result from repeating the route itself, not just any approach to the landmark.

To examine the effect of initially travelling off of the route, but on the habitual side of the landmark, a more distant release site (IV) was used. From there, although the landmark is in a familiar direction, the initial homeward trajectories appear to be guided by PI rather than by route memories: they are straight, rather than curving around the landmark, and they end in search rather than in heading towards the nest (figure 3b). When the ants were then replaced at the feeder, their trajectories showed significantly less confusion than those of the paired ants whose initial trajectories were along the habitual routes from the feeder (p < 0.05, Z = 2.4; figure 3d). Again, ants showed less confusion if their initial trajectories were not along the habitual route.

4. Discussion

The period of confusion observed here on a familiar route has implications both for how the route guidance system works, and for how the route guidance system is coordinated with the other navigational systems [14,24]. Desert ants living in a cluttered environment will generally recapitulate routes reliably if displaced [17–19]. The ants' consistency in such situations can make it appear that route guidance is controlled only by interactions between an individual's current sensory input and its long-term memories, and that such guidance can be modelled simply by a feed-forward neural network [13]. The results described here, however, show that the story is more complicated. First, they show that models of route guidance should in addition allow for regulatory feedback processes that can help control when the route memories are used. Second, they suggest that the relative contributions of the various navigational guidance systems are neither fixed, nor depend solely on the strength of the incoming sensory cues. The interruption of route guidance suggests that guidance systems can be temporarily disengaged so that they do not contribute to the behavioural output from navigational decision-making.

A conceptual model that is outlined in the following sections explains the results in terms of interactions between two regulatory mechanisms within the route guidance system: one is a disengagement mechanism that may be involved in encoding the transitions along a route; the second is a repetition memory that here produces a temporary hiatus where a smooth transition would normally occur. The model presented below may help explain why the landscape structure in this study might produce the patterns of confusion observed here, and why such confusion might not be observed in more cluttered environments [17–19].

(a). Route memories as a collection of guidance elements

Short routes in simple environments can often be explained in terms of guidance using an oriented edge [9,25] or some other simple feature of the visual scene [26]. Ants appear to follow their habitual routes by orienting so that a familiar feature is placed on a remembered position of the retina [11,15]. A retinotopic memory of such a feature (e.g. an edge on a particular position [25] or on a continuous set of positions [11] on the retina) could thus be considered to form a ‘guidance element’. This decomposition of a retinotopic scene into guidance elements contrasts with recent ideas of analysing a scene as a whole [12], but is possibly more compatible with potential attentional processes [27,28] and recent neurophysiology [29]. It also provides a potential way to conceptualize what might be going on in a neural network engaged in route learning [13]. With respect to the current data, it provides a way of explaining why the confusion occurs where it does.

As an ant travels along a natural foraging route, a succession of foreground features will enter, move across, and leave the field of view. Relative motion cues can thus facilitate the selection of suitable features to be learnt and used for guidance elements. Movement can segregate foreground from background [30], and accentuate objects and their edges [31]. Those features close to the direction of travel will generally provide the most reliable cues for guidance as their positions, both on the retina and with respect to the more distant panorama, will remain relatively constant. Once a route has been learnt, an ant's view-based guidance at any moment would be determined by the set of currently active guidance elements. Over some stretches of a route, particularly where prominent features lie either side of the direction of travel or at important goals, more than one guidance element might be active at the same time. In addition, regions of overlap at transitions could ensure that one element (or set of elements) merges seamlessly into the next. Such regions of a trajectory might then reflect an average of the active guidance elements [32]. If no view-based guidance elements are active, and neither are any other guidance systems active, then an ant would engage in search behaviour.

Along the route in the current study, the first 3–5 m could be controlled by a single guidance element based, for example, on the closest edge of the cylinder at a single retinal position (figure 4a,b). Over this section, the edge remains at a relatively constant position, both on the retina and with respect to the more distant panorama. As the ant approaches the cylinder, this initial element would be supplemented and then replaced by other elements that take the ant past the cylinder and along the final approach to the nest (figure 4c). The second element might be based on the same edge, but have a different retinal size and set of positions (see [11]). The observed period of confusion occurs after 2–5 m of normal route following. It could be produced if, in the recapitulation trajectories, an overlap between the first and second elements had been eliminated. The routes would then be resumed once the second guidance element had been activated. What triggers the end of the first guidance element and why might the overlap between guidance elements disappear?

Figure 4.

Using a guidance-element model to explain trajectories along a habitual route. (a) Schematic illustrating how the guidance-element model could explain the period of confusion in the recapitulation trajectories. The three overlapping segments each represent a putative guidance element. Reduced activation during the recapitulations may affect their expression and produce a gap: either ending the first element prematurely, or delaying the initiation of the second. Dotted lines show the direction from the ant to the closest edge of the cylinder. (b) Expression of the first guidance element during recapitulations. Figure shows the same data as figure 1c, but showing in bold the initial segments before any period of confusion. (c) A subset of previously reported zero-vector trajectories [11] suggests that this middle section of the route is encoded, primarily with respect to the features on cylinder itself, as a single segment. In these trajectories, the ants continue too far around the landmark, overshooting the route directions that would be predicted from a more global panorama-matching [12,13], before suddenly breaking off and heading towards the nest. The use of a local integrator in this way could explain why on the current route the ants, which join the route only relatively near the artificial landmark and so would start a local integrator late, then break off from the route memories late, continuing too far around the landmark.

(b). Panoramic context, local integrators and the disengagement of guidance elements

In cluttered environments, a simple feature would generally not be sufficiently distinctive for its detection to ensure, first, that a guidance element is used only at the appropriate location, and second, that it is assigned to the appropriate feature within the visual field. Both problems can be solved, however, if the general panorama provides contextual cues that control the activation of guidance elements (reviewed in [16]). Thus, a guidance element would be active only when situated correctly within the appropriate panorama. It would get switched on when its associated panorama is entered, and switched off when it is left.

In cluttered environments, it might be sufficient for the transitions between heading direction memories to be controlled simply by changes in the current visual input [13]. However, the constraint that transitions in behaviour coincide with transitions in the panorama could be relaxed by incorporating a regulatory mechanism previously proposed to explain learnt routes in bare or homogeneous environments. Honeybees can learn to search for an inconspicuous feeder after travelling a fixed distance within a homogeneous channel [33–35]. It is then the distance travelled within the associated panorama that provides the cue to switch off the route guidance and begin search [34]. Such findings can be incorporated here if each guidance element maintains information about the distance over which it is used in some sort of ‘local integrator’ [34,36]. Each local integrator would be initiated with the activation of its guidance element. If a local integrator exceeds its learnt endpoint before there is a transition in the panorama, then it switches off its guidance element. In this way, transitions in behaviour can occur even in the absence of transitions in the panorama (for instance to search for an inconspicuous feeder). This type of disengagement could also play a role in preventing an ant's path from being captured by a dominating feature (so that it does not behave like a moth circling around a flame, figure 4c). A route guidance system, based in this way on guidance elements that are panorama-specific and that have local integrators to provided stop signals, could operate robustly in both cluttered and open environments.

Uninterrupted route following, for instance in earlier studies [17–19] and in the non-recapitulation trajectories here, can be explained if a guidance element is disengaged only after the subsequent guidance element has already been activated. The confusion observed in the recapitulation trajectories could arise if the repetition memory in some way reduces or eliminates the normal period of overlap. A stop signal from a local integrator can then explain the initiation of the confusion and, importantly, why the region of confusion shifts with the release point (figure 1d).

(c). Short-term memories of visual experience and the repetition effect

What is the repetition memory and how might it reduce an overlap between two guidance elements? One possibility is that the repetition memory arises because the components of a guidance element are only partially reset to their initial states when a guidance element is disengaged—they might be fully reset with a change in motivation (e.g. when the ant leaves the nest or a food site). A guidance element might thus maintain some memory of travel within its associated panorama, and this memory could reduce any subsequent activity level of the element. For example, the repetition memory could be a residual activation in the first element's local integrator so that, during a repetition, the integrator reaches its endpoint faster and switches off the element earlier than normal. Or alternatively, the memory could be a habituation to the associated panorama, delaying the activation of the second element until after the first guidance element has already been disengaged. If the component responsible for the repetition memory were to receive low levels of input when the panorama is a close but not exact match, then such explanations could also account for why low levels of confusion occur even in the non-recapitulation conditions: there could be low levels of input during the off-route initial trajectories (figure 2b,d,f), and similarly, in the extension tests (figure 1d), there would already be some input before an ant reaches the start of the route.

In the experiments reported here, there are a number of factors arising from both the length and the relatively homogeneous visual structure of the approach to the cylinder that might make the first transition particularly vulnerable to disruption. The absence of a sharp transition in the panorama may mean that cues for initiating the second element are weak and, moreover, that a local integrator would be influential in ending the use of the first element. In addition, there might be only a single guidance element either side of the transition, with both elements probably based on the same visual feature and sharing a similar panorama. By contrast, transitions along paths through cluttered environments [18,19] may be less vulnerable as the many objects in the foreground could provide the basis for multiple control elements, and the transitions would occur both frequently and at relatively sharp boundaries in the general panorama.

The effect of repetition on route following observed here suggests that an ant's guidance along a habitual route may involve several classes of memory processes. The visual route memories themselves are long-term memories that are used on subsequent trips that occur days, and possibly even months [37], after the memories are acquired. Some studies suggest that there may also be very short-term memories, lasting no more than 5–10 s [6], that might provide contextual or sequential information [4]. Because the repetition memory revealed here persists over the capture and displacement procedure, which lasts about 30–100 s, it must be a longer-lasting memory of visual experience on the current trip. This memory might well be related to the examples of ‘counting’ that have been reported in honeybees [38]. The guidance-element model illustrates how the repetition memory may arise and why it need bear little resemblance to the general-purpose mammalian episodic memories. Nevertheless, the results show that ants can, in some way, remember that they have recently either performed a specific action or experienced a specific view, and that they can modify their behaviour accordingly.

Funding statement

The BBSRC provided some financial support.