The choreography of learning walks in the Australian jack jumper ant Myrmecia croslandi

Ants, wasps and bees are central place foragers that always return to the nest after outbound journeys. In order to do this, a forager can employ path integration but must also form robust long-term visual memories of the nest and goal locations (reviewed by Zeil, 2012; Collett et al., 2013a). In the case of inexperienced foragers, these memories about the location of the nest or goal location are acquired through a highly structured process of learning during learning walks and learning flights (reviewed by Collett and Zeil, 2018). These visual memories are then used to guide the subsequent approach to the goal.

We are concerned here with the learning walks of pedestrian insects. Namibian desert ants, Ocymyrmex robustior, perform learning walks when confronted with a new landmark around the nest (Müller and Wehner, 2010). The learning walks are spiral-like with well-choreographed rotations along the vertical body axis (Müller and Wehner, 2010). There are short stopping phases (∼150 ms) during these rotations where individuals look back in the direction of the nest. From their perspective, the ants cannot see the nest, so these nest-directed turn-backs must be informed by the current state of their path integrator, and the suggestion is that the views across the nest are remembered at these moments (Graham et al., 2010; Müller and Wehner, 2010). Wood ants Formica rufa also engage in learning walks when they leave a newly discovered feeder and look back at landmarks associated with the feeder and not the goal itself (Nicholson et al., 1999). Backward turns where ants face the landmarks directly become less frequent as the ants become more familiar with the location and with increasing distance from the feeder (Nicholson et al., 1999). It should be noted here that in both published cases, the ants under investigation were already experienced foragers and performed learning in response to a new situation in their familiar environment.

The learning walks and flights around the nest of inexperienced foragers in their natural environment have only recently received more attention. These learning events in ants are characterized by an insect leaving the nest and returning to it repeatedly after walking in different compass directions close to the nest (Wehner et al., 2004; Muser et al., 2005; Stieb et al., 2012; Fleischmann et al., 2016, 2017, 2018a,b; Grob et al., 2017). On a different scale, the ‘exploration flights’ of naïve honeybees (Capaldi and Dyer, 1999; Capaldi et al., 2000; Degen et al., 2015) and bumblebees (Osborne et al., 2013; Woodgate et al., 2016) are very similar, in that successive flights cover different compass directions around the hive.

There are some interesting similarities and differences between the learning walks and the learning flights of insects. First, in both ants and flying insects, learning occurs on leaving the nest for the first time, as the first response to an altered visual environment (e.g. Nicholson et al., 1999; Müller and Wehner, 2010; Narendra and Ramirez-Esquivel, 2017), or whenever the insects had difficulties locating a goal during the previous approach (e.g. van Iersel and van den Assem, 1964; Zeil, 1993a). The more familiar an insect is with the goal location, the shorter and less frequent learning flights and walks become (e.g. Collett and Lehrer, 1993; Lehrer, 1993; Lehrer and Collett, 1994; Zeil et al., 1996; Nicholson et al., 1999; Müller and Wehner, 2010; Robert et al., 2018). In both cases, learning involves moving along arcs and loops around the goal while backing away from it (e.g. Zeil, 1993a; Collett and Lehrer, 1993; Lehrer, 1993; Lehrer and Collett, 1994; Collett, 1995; Nicholson et al., 1999; Müller and Wehner, 2010; Collett et al., 2013b; Philippides et al., 2013; Riabinina et al., 2014; Stürzl et al., 2016). In the case of ants, individuals repeatedly turn back to face the goal (Fleischmann et al., 2017, 2018a), whereas flying insects carefully control where they view the nest as they pivot around it during their learning flights (e.g. Zeil, 1993a; Riabinina et al., 2014; Stürzl et al., 2016).

Here, we concentrate on the detailed spatio-temporal organization of the learning walks of Myrmecia croslandi ants at the nest site, with the aim of understanding how place learning is organized and controlled. Preliminary results have been documented briefly before (Jayatilaka et al., 2013; Jayatilaka, 2014).

Myrmecia croslandi Taylor 1991 (the Australian jack jumper ant; Fig. 1A) are solitary foragers. They show no evidence of relying on recruitment or trail pheromones for finding food: they are seen individually hunting or climbing trees, they move along idiosyncratic paths and orient visually when displaced (Narendra et al., 2013; Jayatilaka et al., 2014). We studied one nest exclusively for data collection for this analysis. The nest is located at the Campus Field Station at the Australian National University in Canberra, Australia (35°16′49.87″S, 149°06′43.74″E). Workers of M. croslandi at this nest predominantly foraged on two Eucalyptus trees approximately 10 m southwest of the nest and hunted for insect prey on both these trees and on the ground (west and northwest of the nest; Fig. 1B,C), which they carry back to the nest (for details of the foraging ecology of this species, see Jayatilaka et al., 2011; Jayatilaka et al., 2014; for details of their navigational abilities, see Narendra et al., 2013; Zeil et al., 2014).

The nest was observed from October 2012 to October 2013. Data on the learning walks of inexperienced and individually identified ants were collected from October 2012 to November 2012. During this period, above-ground activity was high following cessation of foraging activity during the winter months (Jayatilaka et al., 2011). Before above-ground activity had completely resumed, we made regular checks on the nest towards the end of austral winter and beginning of spring (August, September and October), to ensure that recording was started when the nest first became active.

We observed the nest from 07:00 to 19:00 h on consecutive days. Sunrise time in 2012 from 7 October to 30 November varied from 06:32 to 05:42 h. All ants that left the nest (n=74) were individually marked with a water-soluble acrylic paint (Citadel Colours, France) using a four spot, three colour code system (Fig. 1A). Observations were carried out throughout the day, from when the first forager left the nest until no more ants left the nest. No observations were made on rainy and overcast days as pilot studies showed ants to be mostly inactive on overcast days and completely inactive on rainy days (n=5 days when no observations were made owing to unfavourable weather from October 2012 to November 2012). Learning walks were recorded in an area of 40×30 cm around the nest using a Canon HD Legria HFS 10 camera at 25 frames s⁻¹ with an image size of 1920×1280 pixels.

Twelve marked ants out of a total of 74 observed ants were selected at random and used to record learning walks and their complete foraging careers. These ants were marked as soon as they exited the nest for the first time in that season. A complete foraging career encompassed, for each individual, data from the time an individual: (1) first became active above-ground, (2) carried out learning walks, (3) departed the nest to forage for the first time (i.e. after ants had carried out several learning walks and travelled distances over 2 m from the nest) and (4) continued above-ground foraging activity (i.e. continued to forage over the following days after the first foraging trip had been recorded). Although we cannot be certain whether these marked ants had over-wintered, using the methods described above, we were able to ensure that the ants we observed had not been active in the current observation period prior to marking nor foraged in the last 6 months. In addition to video recording (see below), we also noted the daily timing of learning walks relative to sunrise time and determined for each individually marked ant the number of learning walks, the duration of each learning walk and, in cases where learning walks occurred over a single day, the time between learning walks.

The first three to five consecutive foraging trips of each ant were tracked by placing coloured flags behind her as she walked and by recording the flag trail with a differential GPS as described in detail elsewhere (Narendra et al., 2013). Here, we only show the first foraging trip.

We recorded additional learning and homing walks of un-identified and most likely experienced forager ants at the same nest in April 2016 and January 2017 with a higher resolution camera (3840×2160 pixels, Sony FDR-AX100E) at 25 frames s⁻¹ that allowed us to film larger areas of 73×42 cm and 66×38 cm, and with a Panasonic DMC-FZ200 camera (1842×1036 pixels) at 100 frames s⁻¹ to investigate the details of fixation durations.

Video clips were first converted to JPEG or PNG image sequences using Final Cut Pro (Apple, Cupertino, CA, USA), Vegas Pro 13 (Sony Creative Software, Middleton, WI, USA) or QuickTimePro (Apple) and the x–y coordinates of the front of the head and the pronotum (Fig. 1D) were tracked manually frame by frame (40 ms or 10 ms inter-frame interval) with a custom-written MATLAB (MathWorks, Natick, MA, USA) program (Digilite, Robert Parker and Jan Hemmi, The Australian National University). From this, gaze direction could be reliably determined to within ±10 deg (see Fig. 1D, inset).

Using a scale object in the video images, x and y coordinates were converted to centimetres and, after coordinate transformations to make the nest the origin of the coordinate system and aligned with north, the bearing, gaze direction, retinal position of the nest, angular velocity and walking speed were determined using custom-written MATLAB programs (see Fig. 1D for definition of variables). Coordinates and derived variables were smoothed with a 3- or 11-point running average as indicated in the figure legends. Probability densities were determined with the ksdensity MATLAB function using a smoothing window of 9 deg for directional data. To document regularities across ants and learning walks, the probability densities of speed, angular velocity, scanning amplitudes and gaze directions for individual sequences were normalized to maximum, before means and standard errors (s.e.m.) were determined across samples.

There are broadly three kinds of ant movements close to the nest (Fig. 2A,B): the paths of ants engaged in nest excavation (digging paths), the paths of ants heading out to forage and returning from foraging (foraging paths), and learning walk paths. Digging and outgoing foraging paths are more or less straight, with digging paths characterized by immediate, but not necessarily direct, returns to the nest (blue paths in Fig. 2A and left panel of Fig. 2C). Learning walk paths form arcs or loops around the nest with ants entering the nest at the end of the walk (Fig. 2B). These differences are evident in the plots of distance to the nest versus path length that are shown in Fig. 2C and in the inset graphs, which are examples of bearing distributions of three individuals for each of the digging, foraging and learning walk paths. This illustrates that it is only during the learning walks that individual ants cover a wider range of bearings around the nest.

Learning walks are thus characterized by an ant leaving the nest and returning to it after walking in a loop close to the nest (Fig. 2B). As we will see later, however, learning ants move and turn in characteristic ways while moving along these paths, and experienced ants will sometimes also walk in an arc around the nest before heading out on a foraging excursion.

Most learning walks of the 12 identified naïve ants occurred between 2 and 6 h after sunrise (n=30 walks from n=12 ants; Fig. 2D,E), although some ants performed their walks much later in the afternoon, between 8 and 10 h after sunrise (e.g. ants 1, 2 and 5; Fig. 2D). First and second learning walks occurred in a narrow temporal window relative to sunrise always within 2 h of each other, even on separate days, with the exception of those of ant 1 (Fig. 2D, red and blue circles). When learning walks occurred on a single day, the time between them was highly variable, averaging 13.6±33.6 min (mean±s.d., n=12 ants, range 0.5–164 min). Ants carried out between two and seven learning walks (Fig. 2F) over up to 4 days before starting to forage (Fig. 2G), although most ants performed learning walks over 2 days. There was no significant correlation between the duration and the number of learning walks (regression analysis, P=0.28, d.f.=41). Learning walks within the recording area (40×30 cm) had an average duration of 56.8±49.3 s (mean±s.d., n=12 ants, range 12.4–225.9 s). Unfortunately, with the recording equipment available at the time, only a few learning walk paths were confined to the recording area around the nest (Fig. 3), which had to be kept small to be able to determine gaze directions. We estimate that ants covered an approximate maximum distance of 86 cm outside the field of view of the camera (judged by their average walking speed of 1.2 cm s⁻¹ and time of exit from and subsequent entry into the recording area). Tracking of extended learning walks with differential GPS showed that they covered distances of up to 3 m away from the nest (data not shown).

Viewed across 12 ants, there was no consistent direction in which ants carried out their first learning walks (red paths in Fig. 3 and red distributions in Fig. 4) and these first walks were also not necessarily restricted to an area very close to the nest. However, successive learning walks of individual ants occurred in different compass directions around the nest (Fig. 4). Ants differed in the degree to which they covered directions around the nest relative to their first foraging trip (marked by long red line in inset histograms in Fig. 3). However, the group distribution showed no clear, preferred directions between ants (circular mean direction: −99 deg, vector length: 0.0135; Rayleigh z=0.0103; P=1.0). Note that all first foraging trips went south towards the colony's preferred foraging trees (see Fig. 1B), which may have provided guidance with their overhanging canopy.

On a fine spatial and temporal scale, learning walk paths had a meandering structure (Figs 5 and 6) that was caused by very regular scanning movements of the head (gaze in Fig. 5B), which are supported by body movements (Fig. 5C). Ants oscillated between looking in the direction of the nest and in the opposite direction (see light blue line in Fig. 5B). In the example shown in Fig. 5 and in following figures, instances when an ant looked to within ±10 deg of the nest are marked red (looking in the direction of the home vector) and instances when the ants looked to within ±10 deg directly away from the nest (anti-home vector views) are marked dark blue. The ants' scanning movements oscillated between the nest direction (red line and red circles in Fig. 5B) and directions pointing directly away from the nest (dark blue line and dark blue circles in Fig. 5B), because scanning direction was reversed shortly before or after the ant encountered these directions (see oscillating time course of gaze direction; light blue line in Fig. 5B). This is also evident in the regular reversals of the angular gaze velocity shown in Fig. 5D. It is necessary to add one caveat in relation to this detailed analysis: we have to assume that the ant's home vector points at the nest directly, although we do not know when and where exactly the animals start path integrating at the beginning of these walks.

We note the following regularities in this learning walk and the further examples shown in Fig. 6: the ants walked relatively slowly (<3 cm s⁻¹; Fig. 5D) compared with foraging walks (Fig. 6B, top) and their angular velocity profile differed from that during foraging exits (Fig. 6B, bottom). Their scanning movements can be, but do not have to be, very regular, as documented by the periodic auto-correlations of the angular velocity time series for six walks in the left panel of Fig. 6C and the weak or absent periodicities for seven walks in the right panel of Fig. 6C. During part of their walks, however, some ants turned in the direction of the nest at very regular intervals (Fig. S1), encountering regularly spaced nest- and anti-nest directed views (Fig. S1A,B). Successive nest-directed viewing directions changed at a constant rate during such path segments, as documented for six examples shown in Fig. S1C. In general, the reversals of scanning direction were roughly linked to the moments in which the insects were aligned parallel to the home vector so that the ant’s gaze direction oscillated between the direction of the home vector (marked by red circles in Figs 5B, 6A and Fig. S1B) and that pointing away from home (anti-home vector views marked by dark blue circles in Figs 5B, 6A and Fig. S1B). This was the case even during the initial path sections, when the ants predominantly looked away from the nest as they moved away from it (as indicated by their distance from the nest, black line Figs 5 and 6), and during their final return, when they predominantly looked in the nest direction.

If reversals of scanning direction are indeed associated with instances in which the ants are aligned parallel with the home vector, we would expect: (1) a decrease in angular velocity following nest or anti-nest view encounters; (2) scanning amplitudes between reversals of scanning direction to be approximately 180 deg; and (3) gaze directions relative to the nest at the moment of scanning direction reversal to show two peaks, one close to 0 deg (when the ants look in the direction of the nest) and one close to 180 deg (when the ants look away from the nest). We extracted these features from 13 walks for which we have complete paths, and found that the ants' angular velocity was particularly high when encountering nest and anti-nest views and did not significantly decrease following these encounters (Fig. 7A). Because this pattern may be dominated by nest and anti-nest views encountered during initial movements away from the nest and during the ants' return to the nest, we restricted this analysis to sections of walks between the first turn-back to the nest and the moment the ants began their return to the nest (inset Fig. 7A). Again, angular velocity was relatively high during these view encounters and there was no significant decrease of angular velocity following view encounters, most probably because reversals of scanning direction occurred at variable times before and after view encounters (see Figs 5 and 6). For the same reason, the scanning amplitude distribution has a broad peak between 90 and 180 deg (Fig. 7B) and the distribution of gaze directions relative to the nest direction at the moments in which scanning direction is reversed has three peaks at 20, 74 and 141 deg (Fig. 7C). Overall, the distribution of gaze directions relative to the nest broadly and quite uniformly covers all directions between nest and anti-nest views (Fig. 7D).

We also note that the ants did not linger (or ‘fixate’) when they were aligned parallel to the home vector, but rather swept through these alignments, in contrast to what has been described for desert ants (Müller and Wehner, 2010; Fleischmann et al., 2017, 2018a): at the moment of first alignment (time=0 in Fig. 7A), mean angular velocities were close to 150 deg s⁻¹, which is toward the higher end of the angular velocity distribution of learning walks (Fig. 6B, bottom).

High-speed video analysis confirmed that nest-directed views are not associated with prolonged fixations compared with fixations at other times during learning walks (Fig. 7E, lower panel), with the exception of four instances of slightly longer fixations within 10–20 deg of the nest when fixation durations longer than 0.1 s were considered (Fig. 7E, top panel). Finally, the gaze directions of ants during their learning walks were not associated with specific panorama features: Fig. 7F shows the distribution of all gaze directions of ants (black lines) and of gaze directions when looking into the nest direction (red lines) during seven learning walks filmed at 100 frames s⁻¹ and 13 walks filmed at 25 frames s⁻¹ together with the local panorama (top images) and the mean pixel values along vertical pixel columns (blue lines).

We conclude that during their learning walks, the ants experience a systematic sequence of views as they alternate between turning in the nest direction and turning away from the nest at different compass bearings around the nest, and that this sequence must be guided by the state of the ants' path integrator (as shown by Müller and Wehner, 2010), because from their perspective, the ants cannot see the nest entrance. The ants' scanning movements oscillate between the nest and the anti-nest direction despite the fact that the association between the reversals of scanning direction and the ants' alignment parallel to the home vector direction is rather weak overall (Fig. 7C), mainly because the strength of this association varies greatly throughout the execution of a learning walk (e.g. Figs 5, 6 and Fig. S1).

In addition to full learning walk loops, which are terminated by ants entering the nest, we also found that experienced foragers often performed a partial learning walk segment in an arc opposite to their intended foraging direction (Fig. 8A). Other ants left the nest during the same period of observation in much straighter paths (see Fig. 2A). We recorded these walks after several days of experimenting with foragers that were captured and subsequently released close to the nest. These partial re-learning segments are thus presumably triggered by disturbance experienced previously by individuals and exhibit a similar organization as the one we described for full learning walk loops (see bottom right panels Fig. 8A,B). However, in contrast to the learning walks of naïve ants, partial learning walks were immediately followed by a foraging excursion (indicated by large blue arrows in Fig. 8A).

Although it is attractive to suggest that ants memorize the scene around the nest whenever they are aligned parallel to the home vector, looking in (as suggested by previous work, e.g. Müller and Wehner, 2010; Graham et al., 2010; Fleischmann et al., 2017, 2018a) and opposite to the nest direction (as suggested, in addition, by our observations here), a way still has to be found to show that this is the case. One possible avenue would be to identify the navigational decisions made by homing ants, in relation to where they had been during their learning walks (e.g. Fleischmann et al., 2018b). Unfortunately, we did not collect data on the details of the homing paths of identified ants that would allow us to perform this analysis at this stage. However, the behaviour of homing ants in general suggests that such an analysis would be fruitful: homing ants perform regular scanning movements, which are of much smaller amplitude compared with the scanning oscillations during learning walks (Fig. 9), but most importantly, in many cases, they do not approach the nest directly, but walk past it, before correcting their path. It will be interesting to see whether there is a relationship between the navigational corrections of homing ants and the locations they had occupied during learning, in particular those where they had been aligned parallel to the home vector.

Before going out on their first foraging excursion, M. croslandi ants perform two to seven learning walks around the nest over up to 4 days, as has recently also been described for the desert ant Cataglyphis fortis (Fleischmann et al., 2016). Most learning walks occurred between 2 and 6 h after sunrise, with the first two learning walks restricted to a narrow temporal window of 2 h when they occurred on the same day, but also when the first and second walk were performed on different days. Such temporal fidelity in the timing of learning walks may be attributed to the fact that most foragers exit this nest 4–6 h after sunrise (Jayatilaka et al., 2014).

Successive learning walk paths cover different compass bearings around the nest (see also Wehner et al., 2004; Muser et al., 2005; Stieb et al., 2012; Fleischmann et al., 2016), an observation that is strikingly similar to what has been documented for the sequence of learning flights in honeybees (Capaldi and Dyer, 1999; Capaldi et al., 2000; Degen et al., 2015) and bumblebees (Osborne et al., 2013; Woodgate et al., 2016). It is not clear at this stage whether insects choose successive directions randomly or whether they remember where they have been before and head in novel directions.

The learning walks of M. croslandi exhibit a distinct spatio-temporal organization with, at times, very regular scanning movements that lead to a series of systematically changing views toward and away from the nest from the perspective of different compass bearings. We found no evidence that the ants stop and fixate exactly when facing toward or away from the nest, nor that their gaze directions during learning walks are related to dominant visual features in the nest environment (Fig. 7E,F).

Although recent studies of the learning walks in Cataglyphis ants and an older study in Formica were not primarily concerned with investigating the full spatio-temporal dynamics of learning choreography, there are a number of interesting similarities and differences that emerge from a comparison between different ants.

A common observation in ant learning walks is that ants frequently turn back to look in the nest direction, a behavioural element called ‘pirouettes’ by Müller and Wehner (2010) and Fleischmann et al. (2017, 2018a,b). However, in contrast to M. croslandi, the desert ants Ocymyrmex robustior (Müller and Wehner, 2010), Cataglyphis noda and C. aenescence (Fleischmann et al., 2017, 2018a,b) stand still for 130 to 200 ms when facing the nest during these turn-backs, and wood ants walk a few centimetres back towards a feeder (Nicholson et al., 1999) or along a novel route (Graham and Collett, 2006). In addition, both C. noda and C. fortis frequently perform fast 360 deg turns on the spot (Fleischmann et al., 2017). We did observe such complete turns in M. croslandi, but at most two times during a learning walk (e.g. Figs 5 and 6). We document here that turn-backs in M. croslandi alternate with turns in the opposite direction, away from the nest, and it will be interesting to see whether this is also true for desert ants. It is clear, however, that directed turns during learning walks at the nest must be guided by path integration (Müller and Wehner, 2010), because from the ants’ perspective, they cannot see the nest entrance. Interestingly, the desert ant C. noda continues to be able to execute nest-directed turns when celestial compass cues are absent (Grob et al., 2017), but is unable to do so when the magnetic field is disturbed. When the magnetic field is rotated, they turn to ‘virtual nest directions’ as predicted by the extent of the rotation (Fleischmann et al., 2018a).

The learning walks (or exploration runs) of foragers exiting the nest for the first time, as they have now been described in Melophorus bagoti (Muser et al., 2005), C. bicolor (Wehner et al., 2004), C. fortis, C. noda, C. aenescens (Fleischmann et al., 2016, 2017, 2018a,b) and M. croslandi (present study), all end with the ant entering the nest again and therefore are defined as ‘small-scale round trips around the nest entrance that preceded the foraging runs’ (Wehner et al., 2004). The re-learning walks performed by apparently experienced M. croslandi ants after they have encountered disturbance around the nest or had difficulties locating it previously (see also Müller and Wehner, 2010; Jayatilaka, 2014; Narendra and Ramirez-Esquivel, 2017), differ from the learning walks of naïve ants by not ending with re-entry into the nest, but with the ants leaving the nest area to forage (Fig. 8; see also Müller and Wehner, 2010). Re-learning ants pivot around the nest in the direction opposite to the subsequent foraging direction, during which their scanning behaviour closely resembles that of the learning walks of naïve ants. Considering that homing ants often do not approach the nest directly, but make a number of corrective manoeuvres close to the nest, these partial learning walks followed by foraging may supply homing ants with a crescent of nest-directed and opposite views that protect them from getting lost when overshooting the nest location.

What happens during learning walks and what potential navigational information do they provide?

As has been demonstrated before with regard to the learning flights of bees and wasps (e.g. Opfinger, 1931; Becker, 1958; Tinbergen and Kruyt, 1938; Zeil, 1993b), the learning walks of ants are a prerequisite for subsequent homing success with the aid of the landmark panorama (Fleischmann et al., 2016, 2018b). Like M. croslandi, C. fortis ants also perform three to seven learning walks before heading out to forage, and the more learning walks they have performed at the nest that was surrounded by a landmark array, the more closely they searched for the fictive nest position when released at an identical test array (Fleischmann et al., 2017). The trick appears to be the acquisition of nest-directed views from different compass directions, as first suggested by Graham et al. (2010) and modelled by Baddeley et al. (2012) and Dewar et al. (2014). We illustrate the navigational information provided by such views for the specific environment of the nest we studied (Fig. 10A). In addition, we suggest why it may be advantageous for ants to also learn views when facing away from the nest during learning walks (Fig. 10C).

Consider the case of a nest-directed panoramic snapshot taken 20 cm to the north of the nest, facing south (Fig. 10A): the orientation of this snapshot can be recovered by alignment matching up to 6 m away from the nest in this particular environment, because the rotational image difference function (rotIDF; yellow lines in Fig. 10B) comparing the reference view with the view seen 6 m north of the nest has a detectable minimum (bottom image, Fig. 10B). Approaching the nest from the north (dashed line, Fig. 10A), this minimum becomes more pronounced, flanked by increasingly steeper gradients, as the location of the reference image is approached (Fig. 10B). Over some distance, then, nest-directed snapshots provide information on nest direction even at locations an ant may never have visited before (Narendra et al., 2013; Stürzl et al., 2015; Murray and Zeil, 2017). Following that direction, an ant would also potentially know how close to the nest she is if she monitored the minimum of the rotIDF as it progressively decreases (Fig. 10B). The area and the vantage points covered by learning walks thus determine the range over which an animal can detect where it currently is relative to the nest (Narendra et al., 2013; Dewar et al., 2014; Stürzl et al., 2015). This is potentially the reason why Fleischmann et al. (2018b) found that homing by landmark guidance is affected by the space ants had available for performing learning walks. Finally, we note that the effective range of guidance afforded by panoramic views also depends on the particular distribution of objects and on the visual structure of habitats (e.g. Stürzl and Zeil, 2007; Dewar et al., 2014; Murray and Zeil, 2017; Zahedi and Zeil, 2018).

As an ant approaches the nest (schematic black path in Fig. 10C), still under the guidance of the same nest-directed snapshot taken from a position 0.2 m north of the nest (marked red in Fig. 10C), a second snapshot taken 0.2 m south of the nest, but facing away from the nest (marked blue in Fig. 10C), will gain in ‘familiarity’ (sensu Baddeley et al., 2012) as the minimum of its rotIDF becomes smaller (blue curves in right panels, Fig. 10C). The navigational instruction the ant can derive from the northern snapshot when it has reached close to the location from which it was taken during a learning walk is to keep moving in the direction of the minimum of the rotIDF (red curves in left panels Fig. 10C, direction indicated by light blue arrows in the schematic), which is zero at that location, marking the most familiar of all the learning walk snapshots taken at different orientations and bearings. As the ant continues to move in that direction, small lateral displacements may take her past the nest entrance as shown in this example, which causes the southern snapshot to become more familiar. However, the instruction associated with that anti-nest snapshot is to turn 180 deg to face the nest entrance (indicated by light blue arrow attached to blue snapshot location in Fig. 10C). Note that this manoeuvre would align the ant with neighbouring nest-directed snapshots (marked pale orange in Fig. 10C) and therefore would redirect the ant towards the nest (see also the homing paths shown in Fig. 9).

These considerations suggest that there may be at least two reasons why M. croslandi ants (and possibly other species) systematically intersperse views that are nest-directed and views that are directed away from the nest: first, the latter may provide negative examples improving classification, as has been shown in machine learning and image classification (e.g. Kherfi et al., 2003; Baddeley et al., 2012); and second, if these views are more specifically associated with nest direction (not only learnt as negative examples), they can improve homing efficiency by aligning homing ants that are facing away from the nest with nest-directed views without the need for large-amplitude scanning movements. For instance, scanning movements with 180 deg amplitude would be needed to align the example ant shown in Fig. 10C that had overshot the nest entrance with the nest-directed views south of the nest.

The question remains whether ants during their learning walks only store views when aligned parallel to the home vector. For desert ants, this suggestion is motivated by the observation that they stop and fixate when looking in the nest direction (Müller and Wehner, 2010; Fleischmann et al., 2017, 2018a) and for M. croslandi it is motivated by the observation that scanning direction tends to be reversed around the time the ants are aligned in either direction parallel to the home vector. Another hint comes from modelling studies, where the suggestion is made that alignment with and the decreasing length of the home vector could be used as a reinforcement signal for learning views along routes (Ardin et al., 2016; Webb and Wystrach, 2016). However, so far, there is no evidence to reject the possibility that ants learn all the views they experience during their learning walks tagged with the home direction provided by their path integration system. When a homing ant was aligned with any of these views, each would contribute an instruction in which direction to move to pinpoint the nest, as suggested for homing wasps by Stürzl et al. (2016).

Our detailed analysis of learning walks raises a number of intriguing questions regarding the acquisition, storage and use of views for homing. Why is place learning in central place foraging insects so distinctly choreographed in time and space? We suggest that the spatio-temporal pattern in these learning procedures reflects three aspects of navigation that are insufficiently understood: (1) what rules guide the (anticipatory) acquisition of information needed for successful homing; (2) how is the quality of information assessed during acquisition; and (3) considering that the transition from interior to the outside (when ants move from within the nest to the surface to carry out learning walks) triggers changes in the mushroom bodies of their brains (Stieb et al., 2012; Grob et al., 2017), what temporal organization of behaviour is required by the underlying neural processes to form stable memories, both at the time scale of seconds (e.g. Bittner et al., 2017) and at longer time scales for the formation of long-term memories (e.g. Hourcade et al., 2010; Falibene et al., 2015)? It is worth noting in this context that learning walks and learning flights constitute an unusual form of learning, with no obvious reinforcing or unconditional stimulus or associative context (except potentially path integration, see above), but with features resembling perceptual learning or recognition memory, which are thought to be involved in imprinting (e.g. Horn, 1998; McCabe, 2013). Finally, there is a need to investigate whether the systematic scanning between view directions parallel to the home vector is a common feature of ant learning walks, or an idiosyncrasy of M. croslandi. In any case, it raises interesting questions regarding the organization of navigation-relevant memories and leads to the prediction that if ants only memorize views when aligned parallel to the home vector during learning walks, they should execute 180 deg turns when missing the nest entrance during their approach to the nest. A prerequisite for answering these questions and testing these predictions will be a detailed comparison of gaze directions and navigational decisions during the learning walks of identified ants and during their subsequent homing paths, as it has been recently done for the learning and homing flights of honeybees, bumblebees and wasps (Dittmar et al., 2010; Hempel de Ibarra et al., 2009; Collett et al., 2013b; Stürzl et al., 2016).

We thank Chloé Raderschall, Animesh Agrawal and Teresa Iglesias for their help with field work and with data analysis. We are grateful to Hwan-Yin Joon from the Statistical Consulting Unit (SCU) at the ANU for his help with statistical data analysis. We thank Tom Collett, Pauline Fleischmann, Robin Grob, Wolfgang Rössler, Wolfgang Stürzl and Rüdiger Wehner for their comments on an early draft of the manuscript.