Study Site and Species
The field site (23°45′28.12″S, 133°52′59.77″E) was located on the grounds of the Centre for Appropriate Technology campus, ~10 km south of Alice Springs, Northern Territory, Australia. Melophorus bagoti inhabits visually cluttered areas with vegetation (buffel grass and scattered Eucalyptus). Experiments were conducted during the Australian summer (November through February) in 2022–2023.
Foraging arena
A plastic feeder (15 × 15 × 10 cm) was sunk into the ground 7m from the nest entrance and baited with both cookie pieces (Arnott™) and mealworm pieces. The smooth walls of the feeder prohibited escape after food collection, allowing experimenters to manually start the inbound journey by lifting foragers to the feeder ledge. Between the feeder and nest entrance, all vegetation was cleared, and a 15-cm-high plastic wall was erected, forming a ~3m wide arena (Figure 2A,B). This arena extended 5m, where a 15cm high barrier prevented further movement to the feeder except through a funnel comprising two (61cm H × 91.5cm L) white sheets of Masonite board with a 30cm gap between them (Figure 2C,E). This greatly reduced the amount of natural food resources the nest could access, motivating foragers to search the area until finding the feeder (Figure 2D,E).
Experiment 1: Route Formation
All foragers emerging from the nest entrance were marked on the gaster with enamel paint (Tamiya™) for five consecutive days, denoting them as experienced individuals and allowing us to exclude them from testing. On the sixth day, when non-painted naïve individuals first left the nest to complete a pre-foraging learning walk, they were collected and individually marked using a colour combination of paint (Tamiya™) on their gaster and/or thorax. Marked individuals were observed during subsequent learning walks and upon their first foraging trip. The end of learning walks and onset of foraging was determined by the distinct lack of a learning walk with alternating nestward/outward scanning behaviours. Melophorus bagoti exhibits highly structured learning walk scans (akin to pirouettes in Cataglyphis), rhythmically alternating scanning bouts towards and away from the nest with clear nestward and outward fixations (See Supplemental Materials Video 1, Figure 1). Upon starting foraging, their outbound foraging paths to collect food and inbound returns to the nest were recorded via pencil and graph paper. We additionally marked the location along the path when foragers performed scanning behaviours and lookbacks. After the forager collected food, we also recorded the inbound path from the feeder back to the nest, marking the location of inbound scans as well as lookbacks where the forager turned back to the feeder direction.
We defined scanning behaviour as the stopping of forward movement accompanied by rotational body movement on the spot (interrupted by fixations), while remaining oriented outward away from the nest. Lookbacks are structurally somewhat similar to scans (in that they entail a stopping of forward movement and rotational movement interspersed with fixations); we differentiated these from scanning rotations by constricting a ‘lookback’ to only occasions when rotational movement brought the individual’s orientation to within ±90° of the nest (during an outbound lookback) or feeder direction (during an inbound lookback).
For each individual, we collected positional data of the path as well as any scans and lookbacks observed on the first 10 foraging trips conducted on the first day of foraging. To determine if route formation metrics changed after the overnight delay period when the foraging force is not active, we continued to collect these metrics on each forager’s second day of foraging for the first five trips, also noting if each forager conducted any pre-foraging re-learning walks around the nest prior to second day foraging. After completing this 5th next-day trip, ants were marked with paint as tested. Before testing, we determined that ten foraging trips would be sufficient for a forager to become highly experienced of the route (Freas and Cheng 2018; Freas and Spetch 2019). If a forager conducted less than ten foraging trips on the first day, we continued to collect this individual’s metrics on subsequent days until they reached ten total foraging trips and then recorded their first five foraging trips on the following morning (four individuals took two days to complete their first 10 foraging trips). If a forager continued to forage after ten day-one trips, we continued to collect these paths but given these paths were highly similar to Trip 10 by every metric, they have not been used in our analysis.
Route Formation Data Analysis
Pathswere digitized bymarkingevery ~10cm along the path length while both scans and lookbacks counts were recorded where they occurred along the path. We assessed three metrics of path formation for both the outbound and inbound portions of the foraging route: 1.) path straightness, denoting the efficiency of the outbound/inbound journey between the nest and feeder; 2.) outbound/inbound Scan counts, associated with navigational uncertainty; and 3.) outbound/inbound Lookback counts, associated with route formation and learning. Path straightness was calculated by the ratio between the straight-line distance (nest-feeder) with the forager’s path length. All metrics were compared across trips over two days (fixed effects) using poisson loglinear General Linear Mixed Models (GLMMs) for count data and Gaussian loglinear GLMs for path straightness data with individuals as a random effect. When significant effects were uncovered, a priori within-individual contrasts akin to Helmert contrasts (with α set at 0.01 to correct for multiple non-independent comparisons) were conducted between each trip and the mean of the following trips that day (i.e. Trip 1 vs Trips 2-10; Trip 2 vs Trips 3-10, etc.). When the comparisons became insignificant, we denoted this as the within-day performance asymptote. Contrasts analysing the overnight performance decline were assessed by comparing Trip 1 of the next day with the average performance on the final three trips of the first day (Trips 8-10). Next-day performance comparisons were conducted with Helmert-style contrasts in a parallel fashion to the first foraging day (Next Day Trip 1 vs. Next Day Trips 2-5, etc.), with performance asymptote similarly determined (when p > 0.01). Additionally, within individual comparisons between performance on the outbound and inbound routes of the same trip were compared using Wilcoxon Signed Rank Tests.
We separately analysed the effect of the overnight delay on route formation and maintenance by comparing the mean change (Δ) within each individual in each of the three path-based metrics separated into the outbound/inbound routes (path straightness, scans and lookbacks) between trips characterised by whether they followed a trip that occurred on the Same Day (Trips 2–10 and Next Day Trips 2–5) or after the ~16h Overnight Delay (Next Day Trip 1). For the four ants which took two days to complete the first ten trips, their first trip of day two was averaged with the Next Day Trip 1 (foraging day three for these individuals). Within-individual comparisons between Same Day and Overnight Delay trips were compared using Wilcoxon Tests.
Experiment 1: Lookback Positions
Image analysis of route
To assess potential mechanisms underlying when foragers decide to lookback, we chose to characterise how the visual scene changed as foragers moved between the nest and feeder, looking at three factors: ‘local’ panorama change, ‘global’ panorama change, and the ant’s current global vector length (which is not based on views). We quantified the rate of local panorama change by assessing the panorama change over the last 50cm of the route and more globally by comparing visual scenes across the arena to the panorama at the nest entrance. To accomplish this, we collected 360° panoramic images of the foraging route for image analysis. These images were taken at the nest entrance and at 50cm intervals in a grid-like fashion along the arena leading up to and including the feeder panorama, totalling 72 images throughout the foraging route (Figure 2B-D, 6A).
To calculate a rate panoramic change along the foraging route, we compared each panorama image within the arena with both the nest image for a metric of how the panorama has changed globally from the beginning of the foraging trip (nest), as well as comparing each image to the previous image 50cm before along the grid’s straight-line route to the nest. This gives an indication of how much the visual scene has changed recently as the ant travels to the feeder. As a final metric of change, we also calculated the vector distance (cm) from the nest for each image site. Outbound lookback counts were assigned to the closest image site within the arena (within 25cm). This allowed us to assess the panorama characteristics at sites associated with the number of lookback behaviours which occurred in the vicinity throughout the first 15 trips of the foragers’ careers. Lookbacks that occurred within 25cm of the nest entrance were excluded from this analysis as we could not assess the local or global panorama changes under this distance from the start.
To calculate the amount of visual change between each of the image pairs, rotational image difference functions (rotIDFs) were conducted by using the mean pixel difference (MPD) between the focal image and comparison images for all possible rotations of the test images (in one-degree steps) using custom written scripts in MATLAB (for further details, see Zeil et al. 2003, 2014; Stürzl and Zeil 2007). Each image was down sampled to 1 pixel per degree with final images measuring 360 pixels width × 180 pixels height and converted to the blue colour channel only. When two images contain some degree of familiarity, this produces a rotIDF that contains a clear valley of mismatch or low Mean Pixel Difference (MPD) between the two panoramas highly associated with the direction of translation between the initial image and the image of the updated location, indicating a level of similarity the navigator can use to orient.
To calculate a similarity metric corresponding to the degree to which the two panoramas were similar, we took the difference between the minimum and mean MPD of each rotIDF comparison. This created a ‘panorama change’ metric where high values corresponded with a deep valley of mismatch between panoramas, denoting a large degree of similarity and low levels of panorama change across space. In contrast, low values corresponded with a shallow valley of mismatch between panoramas in the rotIDF, denoting a low degree of similarity between the images and more rapid panorama change over space. A multiple linear regression analysis was conducted to examine the effect of the Local Panorama Similarity (last 50cm), Global Panorama Similarity and Vector Length on the location of Lookbacks during the outbound portion of the paths.
Experiment 2: Lookbacks Choreography
After observing individuals during the route formation experiment, we chose two locations along the route at which to record lookback behaviours for gaze direction analysis; the first ~80cm from the nest entrance and at the funnel area at 5–6m from the nest, given these two locations were associated with a high number of lookback counts during route formation testing. At these sites we spread white sand over the red soil over a 1m x 1m area in order to more clearly observe the ant’s body/head positions on video frames. Above each site we positioned a Sony Handycam 70cm above the ground, with the lense facing down (3840 x 2160, 25fps). A forager which had already been to the feeder at least once generally left the nest in the feeder directional quadrant where the camera’s FOV was positioned. As these foragers left the nest they were recorded until they left the camera frame, capturing any lookback behaviours occurring in this area. Once they left the filming area, each forager was collected and marked as tested to prevent refilming on subsequent trips.
Ants’ gaze directions and positional estimates were manually extracted from the videos using SLEAP animal positioning software (Pereira et al. 2022). Two points along the ant were selected to assess gaze direction during lookback behaviours, the centre of the mouth/front of the head and the head base, where the head and body meet. There two points allowed us to assess gaze direction of the animal and directional and positional fixation phases beyond the body’s orientation, essential given the large degree of body contraction we observed during lookbacks (~55° divergence between gaze and body orientations, See Figure 1A fixation 6). We began tracking the animal’s pose estimates five frames before the ceasing of forward movement associated with the start of the lookback scan and pose estimates were stopped after the ant resumed forward travel to the goal (feeder/nest) for five frames.
When assessing the structure of lookback scans, we collected the number, direction and duration of fixations throughout the lookback, determining the longest fixation duration, and its direction in relation both to the direction of travel prior to stopping and the nest/feeder direction. We additionally chose to analyse the direction at which the inflection or reversal point of the turn occurs (when the individual switches from turning left to right).
Lookback Gaze Analysis
Directional gaze-fixation data was analysed using circular statistics. To determine whether the longest gaze fixation and the reversal direction were directed, we first used Rayleigh tests for circular data (Fisher 2993). If headings were directed, we further analysed whether the mean direction of the longest gaze fixation was in the nest/feeder direction and the opposite (180°) of travel using the 95% confidence interval (CI) around the mean of longest gaze fixations. 95% CIs were calculated through the standard error of the mean heading direction based on the mean vector length (Fisher 2993). Within-individual directional comparisons between longest gaze fixation and reversal direction were conducted using Moore’s Paired Tests.
Comparisons between the two outbound conditions, Near Nest, Funnel and the Inbound lookback conditions were conducted using poisson loglinear General Linear Models (GLM) for count data, with individuals as a random effect. If there was a significant effect of condition, Dunn-Bonferroni Post-hoc pairwise comparisons were conducted comparing conditions. Pairwise comparisons within conditions (absolute angular error comparisons and pre vs. post reversal comparisons) were compared using Wilcoxon tests.