Carpenter ants use diverse antennae sampling strategies to track odor trails

Ants exhibit remarkably sophisticated and diverse mechanisms of navigation that make use of both internal and external cues (Knaden and Graham, 2016). Sensory signals for navigation come through many modalities, including sight (Narendra et al., 2017; Graham and Philippides, 2017), touch (Klotz and Reid, 1991; Seidl and Wehner, 2006) and smell (Steck, 2012). Olfaction, in particular, is crucial for the survival and organization of the colony. Ants have highly developed olfactory systems, which they use to communicate with and recognize their colony-mates (Hölldobler and Wilson, 1990; Greene and Gordon, 2003) as well as to locate food sources and their nest (Steck et al., 2011). Ants use species-specific pheromones to create extensive networks of trails that play many roles in colony life, including recruitment of nestmates to active food sources (Traniello, 1977; Hölldobler and Wilson, 1990; Morgan, 2009; Czaczkes et al., 2015). Studies have shown how such trails are generated and adapted over time (Reid et al., 2010; Czaczkes and Heinze, 2015; Fonio et al., 2016), and even how they can encode polarity towards or away from the nest site (Jackson, 2004). However, we know very little about how ants are able to accurately track these trails.

Previous studies involving crossing or amputation of antennae have suggested that ants use bilateral odor comparisons (tropotaxis) to follow trails (Hangartner, 1967; Schone and Strausfeld, 1984). Theoretical tropotaxis models predict that ants turn towards the antenna with the higher odor concentration while following a trail (Calenbuhr and Deneubourg, 1992). A more recent study looking at ant behavior at a trail bifurcation point has shown that the angle of turning is proportional to the difference in the total quantity of estimated odor on either side of the bifurcation (Perna et al., 2012). This suggests that ants can discriminate small differences in sensory input to the two antennae and these differences influence motor output. Despite the evidence of bilateral antennae comparisons, little attention has been given to the actual behavioral strategies and the patterns of antennae movements used to actively sample physical space during tracking.

To address this gap in knowledge, we designed a behavioral assay that used high spatial and temporal resolution videography to quantify how ants use their antennae while tracking odor trails. We used the black carpenter ant, Camponotus pennsylvanicus (De Geer 1773), which is widely available and easy to maintain in the laboratory, and which exhibits robust, innate trail-tracking behavior. We found that tracking behaviors are much more complex and diverse than previously described. Moreover, antennae usage revealed many interesting properties of insect behavior including stereotyped motor patterns, differences in left–right usage and behavioral plasticity after amputation. Thus, scent-tracking behavior in ants represents a rich model for understanding sensorimotor transformations and for quantitative biology in general.

Several colony fragments (∼50–100 individuals without a queen) were collected in the Middlesex Fells Reservation in Medford, MA, USA, during spring and summer months. Individual colony fragments were maintained separately in 17×12×7.8 inch (Hefty High Rise) plastic containers with sides coated with Fluon (BioQuip, Rancho Dominguez, CA, USA). Containers were lined with leaf-litter and small pieces of dead wood from the collection site. A glass test tube (150×25 mm, VWR) was humidified by adding ∼10 ml of water and plugged with cotton balls. This tube was placed in the container as a nest area and covered in aluminium foil to block out light. Ants were given fresh water and food (Bhatkar and Whitcomb, 1970) twice per week and kept at constant temperature (23°C, 20% humidity) on a 12 h light:dark cycle (Sylvania Gro-Lux 20 W, 24 inch lights).

Ants were first anesthetized by placing them in the −20°C freezer for 5 min. Ant gasters were surgically removed with a razor blade; 250 mg of gasters were put into 0.5 ml 70% ethanol and crushed vigorously with a blunt metal rod. The resulting extract was filtered through a 0.45 µm pore filter (VWR sterile syringe filter). Blue fountain pen ink was added (10 µl) for visibility. Ink alone in ethanol did not elicit a tracking response in ants. This liquid was stored at 4°C and kept for several months.

A circular platform (1 m diameter) with 150 mm plastic walls (coated with Fluon) was used as an arena for imaging trail-tracking behavior. An enclosure measuring 1 m³ was built around the platform to block light and air currents. Three IR light sources (Phenas, 48 LED array 850 nm) were used to illuminate the imaging area. A camera was mounted at the top of the box (DMK 23UM021, 1280×960 pixels, The Imaging Source) to image a 28×21 cm area at 0.22 mm per pixel. The bottom of the arena was lined with fresh butcher block paper (Boardwalk, 36 inches×72 feet) before each imaging session. A piece of letter-sized index paper (Limited Papers, 110 lb stock) was taped onto the butcher paper below the field of view of the camera. On this paper, a light pencil trace of the trail (15 cm long) was drawn. Next, the trail pheromone was laid down by evenly pipetting the trail extract along the pre-drawn pencil trace with a 10 µl pipet.

An ant was captured at random from a colony and put into a small acrylic box (15×13×6 cm). This box was then placed in the imaging arena on top of the stock paper. The box had a small opening (2×0.5 cm) on one side, which was positioned at the start of the trail. As the ant explored the box, it would find the opening and encounter the trail (this would take 1–2 min). This step was necessary to calm and direct the ant, so that it did not immediately run away from the release site and fail to encounter the start of the trail.

Images were acquired at 60 frames s⁻¹ using IC Capture 2.3 (The Imaging Source) and videos were processed using ImageJ (Schneider et al., 2012). Videos were down sampled to 30 frames s⁻¹ for analysis. We stopped recording after the ant tracked the entire trail or abandoned the trail completely. Each ant was given one trial and then returned to the colony. Each trail was used to run multiple ants for up to 1 h.

All analysis was carried out in MATLAB (MathWorks, Inc.).

To extract trail coordinates, pictures of the trails were taken under visible light before any ants were introduced to the trails. From this image, pixels were classified as part of the trail using an adaptive threshold function in MATLAB. A uniform odor concentration was assumed across all pixels classified as part of the trail.

For each video, a custom-written MATLAB script was used to quantify specific parameters describing the ant in each frame, including the center of mass, the center of the head, the XY positions of the tips of the left and right antennae (absolute and relative to the center of the head), and the body angle of the ant (Movie 1). From these variables, speed, turning angle and trail overlap of each antenna were derived. Software can be found at: https://github.com/rwdraft/AnTipMarker. For future studies, we intend to use a fully automated markerless tracking method based on deep neural networks (Mathis et al., 2018).

Heat maps were generated by plotting antennae tip position relative to the head position in each frame. To pool data across differently sized ants, data were normalized by antenna length in both X and Y directions. Antenna length in each dimension was taken to be the farthest X and Y distance, respectively, from the head to the antenna tip over the entire run for each ant. Normalized relative position was displayed as a 2D distribution over 50 bins in both X and Y directions.

From the tracking videos, three distinct patterns of behavior (‘probing’, ‘sinusoidal’ and ‘trail following’) were identified through observations of video recordings. An algorithm was created that used speed and the distance from the head to the trail to segment videos into these different behaviors that match manual annotations made by several researchers (Fig. S1). First, if an ant's speed was less than or equal to 2 pixels per frame (13.2 mm s⁻¹) for five consecutive frames, the ant was said to be stationary and these times were marked as probing. Second, sinusoidal movement was identified by examining the distance of an ant's head to the trail over time. To ensure this distance scaled with body size, which varies among ants, a relative measure was used as a base unit: a ‘segment’ is the distance from the center of the ant's head to its centroid (∼4.0 mm). The presence of any prominent and distinct peaks in the distance between the ant head and the trail over time was noted (i.e. peaks greater than one-sixth of a segment and not closer than two-thirds of a segment to each other). To characterize the behavior as sinusoidal, there had to be at least two marked peaks within 1 s of each other (i.e. periodic movement about the trail). Third, trail following was identified as times when the distance from the ant's head to the trail was less than one-third of a segment for at least five frames. Additionally, ‘off-trail’ behavior was identified as times when the distance from an ant's head to the trail was greater than two-thirds of a segment, and any unmarked frames were characterized as ‘other’ behavior.

A rough estimate of the odor concentration sensed by the antenna was obtained by assuming (1) the antennae are held close to the surface (i.e. ignoring Z-dimensional information), (2) the antenna's ability to detect trail odor is approximately a step function of distance, (3) the trail pheromone concentration is uniform among pixels labeled as being part of the trail, and (4) the majority of odorant receptors are at the tip of the antenna (Nakanishi et al., 2009). A circle of 10.5 pixel diameter centered on the antenna tip was drawn. The number of trail pixels within this circle is the measure of ‘trail overlap’ or estimated odor concentration (Movie 2). To explore how the width of the step function (the circle) affected our estimation, this trail overlap was plotted as a function of diameter of the circle. Empirically, a circle of diameter 10.5 pixels (2.3 mm) preserved the most information without saturation or redundancy (Fig. S2) and was used for all analyses. For reference, the trail width was approximately 10 pixels (2.0 mm).

A vector was calculated for the movement of the antenna tip between frames. To prevent alignment imperfections from contributing to the data, data points were discarded if the Euclidean distance of the movement between frames was less than 4 pixels (0.88 mm). A histogram of the movement angles for each antenna was plotted. The angles from the right antenna were flipped horizontally to match the perspective of the left antennae so data from each behavioral module could be pooled into a single plot.

To quantify the angle between the flagellar segment and the scape of the antenna, the position of the bend between these two segments was marked manually for both tracking and freely exploring ants. Two lines were drawn between the three marker points: (1) between the antenna tip and the joint and (2) between the joint and the head point. The angle between the two lines was calculated for each frame for both antennae.

To quantify how the antennae interact with the trail during probing and trail-following behavior, the duration an antenna tip stayed on the trail once it made contact was measured as well as the maximum trail overlap it had during contact. This was done in two different ways, which gave similar results. In the first method, the ‘findpeaks’ function in MATLAB was used on the trail overlap data. Any peaks closer than 10 frames to each other were discarded, and the width of each peak was measured at the half-height of the peak. This method removes data from nearby peaks and therefore may underrepresent the duration of continuous encounters with the trail. In the second method, a modified ‘findpeaks’ function was used. The extent of all peaks at the half-height was recorded. Peaks with overlapping extents were considered a single peak and the highest peak value was taken as the true peak height. The longest extent was taken as the true peak width.

For trail following, a linear regression was performed on the change in an ant's body angle using the distance of each antenna tip from the trail as a predictor. Distance was used instead of our estimated odor measurement because distance is a more continuous variable. Data from frames where both antennae tips were far from the trail (i.e. both had 0 values of trail overlap) were removed. The body angle change at different time points was regressed with respect to antennae tip position (0–8 frames ahead in time; data not shown) to identify any lag between behavioral output and odor input. The highest coefficient of determination (adjusted R²) value was found by pairing body angle data from 4 frames after tip position data. In no case did ant speed improve the model as a predictor of body angle change, so it was not included in our final regression. In Fig. S4A, this change in body angle was cross-correlated with the change in the angle of the antenna relative to the head. The antenna angle was calculated by taking a line from the center of the head to the tip of an antenna and finding the angle between this line and the body axis. A negative change in antenna angle is a movement towards the body axis and vice versa.

The distances from the left and right antennae tips to the trail (or to the midline) were correlated during ‘bouts’, or sequences of frames for a single behavior of at least 30 frames (1 s). To determine significance, correlation values were first converted using Fisher's Z-transformation.

To model the expected difference in the left and right antennae trail overlap, two methods were used. In the first, the trail overlap values (right and left) for a given ant were pooled. Then, half of these values were randomly assigned to one antenna and the remainder to the other antenna, and then a bias measure was calculated: (sum of left values−sum of right values)/sum of all values. This was repeated 100,000 times to generate a bootstrap distribution. Only ants with more than 200 frames of data were included in the analysis.

The second method took into account the total number of peaks in the trail overlap data, which was obtained using the modified ‘findpeaks’ function described above (see ‘Odor peak height and width’). The total trail overlap values (right and left) for a given ant were summed and this number was divided evenly into the number of peaks present in the observed data. Each peak was then randomly assigned to either the left or right antenna and a bias measure was calculated as described above. This was repeated 100,000 times to generate a bootstrap distribution. Only ants with more than 200 frames of data were included in the analysis.

For antenna removal experiments, ants were captured from a colony and anesthetized by placing them on a bed of ice (covered by a wet paper towel) until they were immobile. Under a stereomicroscope, one antenna was clipped off at the base of the head using sterile scissors. These ants were then marked for easy identification by applying a spot of nail polish to the abdomen with a toothpick. These ants were tested in the behavioral arena between 2 and 72 h after surgery.

Larger (and inherently longer duration) turns were identified by finding peaks of greater than 15 deg prominence in the body angle data over time for each ant (using the ‘findpeaks’ function in MATLAB).

In this study, we focused our analysis primarily on straight-line trails where all sensory information apart from olfactory cues from the trail (visual and tactile cues) were minimized.

To study the differences in antennae usage during trail tracking, we compared the antennae positions (relative to the head) for ants introduced to an odorant trail (‘trail tracking’; Movie 1) with those of ants not given a trail and allowed to run freely in the behavioral arena (‘exploratory behavior’). We examined relative positions to focus on antennae sampling independent of head movements. We focused on the antennae tips (Fig. 1A), because this is the region of the antennae where odorant receptors are most highly concentrated (Nakanishi et al., 2009). When tracking odor cues, ants use their antennae to sample a large, arc-like region of space around the midline (where the trail is typically located) (Fig. 1B; Movie 2). During exploratory behavior, the antennae tips are held within a smaller area located away from the ant's midline. Although the mean speed of the ant centroid was significantly lower during trail tracking relative to the exploratory behavior (Fig. 1C, left), the mean speed of antennae tip movement (relative to the head) was statistically equivalent (Fig. 1C, right). Therefore, small antennae movements occur with similar speeds in the two conditions, which shows that ants move their antennae actively whether or not they are tracking an odor trail. However, while tracking, the small movements range over a much wider space including the area directly in front of the ant.

While tracking, ants exhibit substantial differences in their body movements and apparent behavioral states (Fig. 2). Ants moved in sinusoidal paths centered near the trail (sinusoidal behavior), moved at very slow speeds along the trail (probing behavior) or followed the trail very closely and accurately (trail-following behavior). All ant trials included one or more of these three behavioral modules (Fig. S3A,B), which were codified using empirically derived parameters (see Materials and Methods). We marked when ants were far from the trail (‘off trail’) and when ants exhibited behaviors that did not conform to the above categorizations (‘other behavior’). The presence of these easily observable and classifiable behavioral states among a population of ants demonstrates that trail tracking involves multiple strategies and is not a homogeneous behavior.

We examined the antennae tip positions (relative to the head) separately for each behavioral module and found robust differences in sampling space (Fig. 3A). Antennae tips were close to each other and close to the head during probing behavior. In contrast, during exploratory and sinusoidal behaviors, the antennae were farther away both from the head and from each other. In trail following, the antennae spacing was intermediate relative to other behaviors, and antennae were largely excluded from an area at the midline approximately the width of the trail (∼2.0 mm). To quantify these differences, we examined the distance of an antenna tip to the head (ρ) and the angle of the antenna relative to the ant's body axis (θ) (Fig. 3B). Antennae tips were closer to the head and had more acute angles relative to the body axis for probing than for other behavioral modules (Fig. 3C). Trail following had intermediate values of these parameters, and sinusoidal and exploratory behaviors had the largest values. For both ρ and θ values, all differences were significant except for the comparison between sinusoidal and exploratory behaviors.

There is currently no method to measure the actual concentrations of specific chemicals sampled by the antennae. Therefore, to relate the differences in the sampling space to odor sensation, we measured the extent of overlap between a 10.5 pixel diameter circle centered on an antenna tip with the trail (Fig. 3D). Even though the ant was very close to the trail in both probing and trail following, sampling during probing resulted in greater antennae overlap with the trail (Fig. 3E). Antennae overlap with the trail was greater for both probing and trail following compared with sinusoidal, where the ant was often farther than one antenna length from the trail (Fig. 3E). This shows that differences in sampling among behavioral modules are likely to result in differences in odor sensation.

To better understand the antennae movements that underlie these sampling differences, we investigated the change in position of antennae tips (relative to the head) between subsequent frames (Fig. 4A). These instantaneous movement vectors exhibited three different distributions (Fig. 4B). During probing, ants swept their antennae along a roughly 45 deg diagonal towards and away from the head. In exploratory and sinusoidal behaviors, antennae movements occurred along a roughly 120 deg diagonal that was approximately perpendicular to the movement patterns during probing. During trail following, the antennae moved in a direction perpendicular to the ant's body axis. Investigation of the sequence of antennae movement vectors confirmed that the antennae tips moved back and forth between the two peaks shown in the polar histograms across the different behavioral modules (Fig. S4B).

The antennae contain an elbow-like joint between the scape and the flagellum, which changes its angle to help control the position of the antenna tip (Fig. 4C). We measured the distribution of joint angles in each of the behavioral modules to fully describe the movement of the antennae. Again, we saw three broad groups (Fig. 4D). Probing displayed the widest range of angles (interquartile range: 139–85 deg), indicating swinging of the flagellar segment inwards and outwards from the head. Exploratory and sinusoidal behaviors have the smallest range of joint angles, concentrated at the largest angles (interquartile range: 176–160 deg), indicating a rigid, outstretched antenna. Intermediate values of antennae angles were observed during trail following (interquartile range: 170–135 deg). Taken together, these data suggest a model of how the antennae move during these different behavioral modules, which is summarized in Fig. 4E. These findings show that behavioral modules classified with body movement parameters are associated with different strategies for antennae usage.

To study the bilateral use of antennae during tracking and to better understand how antenna signals affect the body position, we focused on the trail-following behavioral module, when the ant tracks the trail closely and accurately. During trail following, the ant largely maintains the odor trail between its two antennae, as seen on the representative plot of antennae tip positions over time (Fig. 5A). To quantify the duration and degree of antennal tip contact with the trail, we measured the peaks and widths of trail overlap and compared them with those during probing (Fig. 5B). As shown, the trail-following odor peak widths, which represent the amount of time an antenna tip spent on the trail, were much shorter than those during probing (Fig. 5C). The peak heights, which represent the maximal extent of overlap between the trail and an antenna tip, were also higher during probing (Fig. 5D). The smaller and shorter peaks in trail overlap during trail following are consistent with the idea that the ants are briefly bringing their antennae to the edge of the trail and then moving them away.

We created a linear model to examine the relationship between odor sensation and changes in ant movement. Specifically, this model relates the antennae tip positions during trail following and the body angle change 4 frames later. This amount of delay (133 ms) was found to yield the best fit based on the adjusted R² value. In this regression, the explanatory variables are the distances from the two antenna tips to the trail separately, excluding frames where both antenna tips have zero overlap with the trail (i.e. zero estimated odor concentration). Under these conditions, we obtained the relationship: Angle change=−0.27+(0.086×Right distance)−(0.060×Left distance), where both independent variables were necessary and the model was significant. When plotting the differences between the left and right antennae tip distances, we found that the largest changes in angle occurred when bilateral signaling was heavily biased towards one antenna. Based on the way we have defined the body angle, a positive change corresponds to a turn to the left. Thus, when the left antenna comes closer to the trail, the ant turns to the left, and a complementary relationship holds for the right antenna (Fig. 5E). This would result in a movement that keeps the trail centered between antennae.

The above data indicate that the two antennae are frequently at different distances from the odor trail, presumably conveying distinct odor concentrations. To determine whether the movements of the two antennae were coordinated, we calculated a correlation value between the distances of the two antennae from the trail for probing and trail following, the only behavioral modules where the ant is close to the trail. Antennae positions were significantly anti-correlated in both behaviors (Fig. 6A,B). The same result also holds when using the trail overlap variable instead of the distances of the antennae to the trail (Movie 2, Fig. S5). Furthermore, correlations between the left and right antennae distances to the trail at different frame lags revealed a periodic relationship between the two antennae (Fig. S5). This result reflects the periodic nature of the raw distance time series (Fig. 6A), which resembles two sine waves with different phases. To evaluate whether this anti-correlation exists in other behavioral modules, we repeated this analysis for distances of the two antennae from the body axis of the ant. Antennae tip positions relative to body axis were significantly anti-correlated during all behavioral modules except trail following (Fig. 6B). Trail following showed much lower anti-correlation values when using antennae tip to body axis distances compared with tip to trail distances. This suggests that, in trail following, alternating movements of the body axis shift the antennae tips towards and away from the trail (Movie 3). This is opposite to other behavioral modules, where the antennae themselves alternate about the body axis. Overall, across behavioral modules, movement of the two antennae is anti-correlated regardless of how and whether an ant is tracking an odor trail.

Additionally, we observed that one antenna within a pair often overlaps significantly more with the odor trail than would be expected by chance (i.e. random permutation of the same data, example raw data in Fig. 6C). We defined a bias index (i.e. relative difference of trail overlap values, Fig. 6D) for each ant run. The population of ants analyzed showed a continuum of bias from very strongly left biased to very strongly right biased (Fig. 6D, 8 of 22 left biased, 7 of 22 right biased). Since odor is encountered in discrete events when the antenna comes in contact with the trail, we also tested whether data were significantly different from expected if the total odor were randomly distributed in blocks equal to the number of peaks present in the trail overlap data (Fig. S6A, 9 of 22 biased). From the features of the peaks in the trail overlap data, it appears the bias arises from differences in the number of peaks, peak widths and peak heights between the left and right antenna, but each of these factors contributes differently in those ants which show a bias (Fig. S6B).

This continuum of bias, found in straight-line trail-tracking ants, was not present in ants tracking trails of various degrees of curvature (from radius 3 cm to 6 cm, example in Fig. 6E). In curved trails, the biased antennae were nearly always ipsilateral to the inner curvature of the trail (Fig. 6F, 24 of 29 ants significantly ipsilaterally biased, 1 of 29 significantly contralaterally biased). These data suggest that there are biases in left–right usage among ants, but these biases may be context dependent.

Taken together, this shows that the left and right antennae are used in a non-redundant manner during tracking and that ants take advantage of bilateral symmetry during tracking behaviors to gather information about the environment.

To test the importance of bilateral antennae information for tracking behavior, we removed a single antenna from ants and tested their ability to track trails (2–72 h after surgery, n=6 ants, 14 trials). We compared the single antenna of these ants with the left antenna of our previously analyzed tracking ants. The single-antenna ants showed a wider distribution of relative antenna tip positions, including positions farther away from the head and closer to the midline compared with the distribution of control ants (Fig. 7A,B). Furthermore, the single antennae often extended across the midline, which was much less apparent in control ants (Fig. 7A, red pixels). To quantify this difference, we compared the percent of time when an ant's antenna crossed the midline and found single antennae crossed ∼3-fold more often (Fig. 7C). This raised the question of whether these different distributions impacted trail sampling. We focused this analysis on trail following, when the ant was closely tracking the trail, and quantified the percent of time an antenna extended beyond the contralateral side of the trail (Fig. 7C). The single antennae crossed the trail ∼7-fold more often. This shows that single-antenna ants adapt to amputation by changing the pattern of usage of the remaining antenna to sample a larger physical space around themselves and around the trail.

Next, we evaluated how accurately the single-antenna ants were able to follow the odor trail by plotting the distance of the head to the trail over time. We compared the root mean square error (RMSE) of this distance and found that single-antenna ants had significantly higher RMSE (Fig. 7D,F). The side on which errors occurred was not consistently correlated with the position of the removed antenna (data not shown). Additionally, we found that ants with a single antenna made fewer small instantaneous body angle changes relative to control ants as well as a greater number of large (longer duration) turns (Fig. 7E).

Finally, we compared the trail overlap from the single antenna and control ants (Fig. 7G, left and middle boxplots). Surprisingly, we found these estimated odor distributions to be similar, suggesting that the wider sampling area of single antennae helps to compensate for larger head to trail deviation. However, as shown previously, single-antenna ants track less accurately (Fig. 7F); given that the two antennae are used non-redundantly (Fig. 6), control ants receive much more odor information over time. Therefore, we selected, from control ants, the larger of the two antennae trail overlap values at all time points as a comparison (Fig. 7G, right boxplot). Indeed, single-antenna ants showed far less overlap with the trail compared with this bilateral signal. Therefore, having two antennae allows the ant to be in contact with the trail more continuously.

In this study, we quantified and modeled the strategies of tracking behavior and antennae usage in ants (Camponotus pennsylvanicus). We have shown, for the first time, that ants use a variety of behavioral strategies during trail tracking, each of which has a distinct and stereotyped pattern of antennae sampling. In all trail-tracking behaviors, we found that the left and right antenna play non-redundant roles in odor detection: they generally move in anti-correlated fashion and ants often exhibit a bias in which one antenna is closer to the odor trail. This observation suggests lateral specialization and, moreover, the use of complex strategies for odor detection. We further examined the influence of bilateral comparisons over an extended trail-tracking episode, and we found that ants show signatures of edge detection and use information from both antennae to guide turning and accurately follow the trail. Removal of one antenna impairs tracking but also results in adaptation of antenna sampling, suggesting ants can adopt motor strategies to partially compensate for antenna loss.

Quantitative studies have led to a greater understanding of olfactory behavior in many species (e.g. Gomez-Marin et al., 2011; Riffell et al., 2014; Khan et al., 2012). Although chemical communication and trail following in ants have been studied for many decades, there is very little known about sensorimotor strategies at the level of the antenna. Ants have relatively long sensors that are highly motile. As a first step in being able to model their olfactory behavior, we looked at antennae movement during trail tracking at high spatial (0.22 mm per pixel) and temporal resolution (30 frames s⁻¹).

We found that antennae are highly dynamic whether tracking a trail or exploring (Fig. 1C). However, during tracking, antennae sample a much larger area of space (Fig. 1B). Additionally, by measuring ant speed and body movement relative to the trail, we were able to classify behavior into three discrete components (or behavioral modules, Fig. 2) – sinusoidal, probing and trail following – similar to classifications done in other invertebrates (Green et al., 1983). These three behavioral modules showed distinct patterns of antennae sampling.

The first behavioral module, sinusoidal, was characterized by periodic movement along the odor trail. This behavior did not occur frequently in no-trail conditions (1 of 10 ants exhibited sinusoidal behavior). This type of movement has been observed in ants (Hölldobler and Wilson, 1990) as well as in a range of other species (Martin, 1965; Borst and Heisenberg, 1982; Gomez-Marin et al., 2010; Khan et al., 2012; Porter et al., 2007). It is thought to represent a search pattern along an odor trail or gradient, where temporal odor information is used to stay centered on a signal. Although the ant head and centroid cross the trail during this behavior, the antennae do not cross the trail often. Instead, the antennae touch the trail and the ant turns away (as seen in the example plot shown in Fig. 2). Furthermore, we noted a stereotyped movement pattern when correlating changes in antennae- and body-based movement (Fig. S4A); for example, movement of the left antenna towards the body is often followed by a turn to the right. These observations highlight sinusoidal behavior as a more stereotyped movement. In our analysis, we found that the relative antennae tip positions and movements in this behavioral module were restricted to a smaller area (when compared with probing or trail following) and were indistinguishable from antennae positioning during exploratory behavior (Fig. 3B). Therefore, odor sampling during sinusoidal behavior is accomplished by alternating movements of the ant itself along the trail while keeping its antennae outstretched. This is the only behavior where antenna movements show a significant correlation with the movements of the body (Fig. S4).

The second behavioral module, probing, described ants that were stationary and within sampling distance of the trail. During probing behavior, the ant's antennae had the largest area of sampling (Fig. 3B) and were most densely focused on the trail itself (Fig. 5A). Additionally, ants swept their antennae back and forth towards the head (Fig. 4B). These observations are consistent with an intensive sampling of odor to examine and identify its relevance. This behavioral module is not exclusive to encountering the experimenter-drawn trail; ants not given a trail would infrequently engage in a probing-like behavior as well (data not shown).

The third behavioral module, trail following, was characterized by an ant following the trail with high accuracy. The relative antennae tip distribution largely excluded an area at the midline that was approximately the width of the trail (Fig. 3A). Antennae moved in a direction perpendicular to the trail (Fig. 4B), and the short duration and extent of antennae tip–trail overlap results in a higher temporal gradient of estimated odor concentration than in probing (Fig. 5C,D). Taken together, these findings suggest this behavior involves edge detection. Evidence of edge detection during tracking has been seen in other animals such as bats and rats (Yovel et al., 2010; Khan et al., 2012).

In these three behavioral modules, the relative antennae tip positions displayed distinct spatial distributions, and as such they would gather and convey to the brain different information about trail odor (Fig. 3D). Previous studies in various species have analyzed large-scale antennal movements in response to stimuli (e.g. Huston et al., 2015; Pribbenow and Erber, 1996), but finer movements and distinct strategies are largely uncharacterized. Given the importance of understanding how sensors are moved to sample the environment and gather information to guide behavior, our models in ants provide an important basis for future studies.

Previous experimental and theoretical studies have proposed various models of tropotaxis to explain how ants track odor trails (Hangartner, 1967; Calenbuhr and Deneubourg, 1992; Couzin and Franks, 2003). In Argentine ants, behavior at trail bifurcation points could be predicted based on the estimated concentration difference between antennae (Perna et al., 2012). We tested this idea in trail-following behavior by using the distance of the trail to the antenna as a proxy for odor signal. In a linear model, we found that the direction and magnitude of the turning angle of the ant could be partially explained by the difference in trail distance between antennae (Fig. 5E). We found that larger differences in estimated odor concentrations between the two antennae led to larger turning angles towards the trail (Fig. 5E). The best model used a 133 ms (4 frame) delay between odor sensation and subsequent turns. This latency for motor implementation is similar in time scale to that for adult Drosophila, which have been shown to redirect flight heading within approximately 200 ms of detecting an odor plume (van Breugel and Dickinson, 2014). These results demonstrate that ants use bilateral differences in antenna signals to track odors during trail following. We note, however, that putative antennae concentration differences can only partially predict ant movement, and many other factors are likely to be involved in motor decision making.

Given the strong evidence for the importance of bilateral comparisons, we asked whether each antenna within a pair is used differently during trail tracking. We looked at the correlation between antennae using their distance either to the trail or to the ant's body axis. In all behavioral modules, the antennae exhibited a strong anti-correlation (Fig. 6B). Assuming odor information is typically highly variable in time and space, keeping the two odor sensors in different locations would increase the efficiency of detecting odors and sampling odor space. As far as we know, this is the first evidence of stereotyped coordination between antennae during odor tracking. Except for trail following, these anti-correlations arise from alternating movements of the antennae with respect to the body axis. In trail following, antennae predominantly show anti-correlation with respect to the trail but not to the body axis. One possibility is that the antennae movement driving the anti-correlation with respect to the trail comes from changes in the angle of the body axis rather than movements of antennae relative to the body axis. Indeed, careful inspection of the videos shows numerous examples of the ant making small adjustments in body angle with respect to the trail (Movie 3). Thus, like sinusoidal behavior, the body movement of the ant is an important variable involved in positioning the olfactory sensors.

In addition to the anti-correlation between antennae, we found a continuum of antennae bias in trail sampling. In a large fraction of ants (15 out of 22), one antenna overlapped significantly more with the trail than the other during a tracking run (Fig. 6C). In some ants, the left antenna was strongly biased (8 of 15) and in others the right antenna was strongly biased (7 of 15), similar to studies showing a continuum of handedness in turning bias in flies (Buchanan et al., 2014). This bias was radically altered when the odor trails were curved as opposed to straight (Fig. 6E,F). When following curved trails (either to the right or the left), the bias was strongly towards the antenna ipsilateral to the inner curvature. This result could be interpreted in two ways: (1) any inherent bias in sampling (i.e. ‘handedness’) can be masked by trail features that impose a greater bias in antennae overlap; (2) the pattern of antennae sampling used by ants is not inherent but rather context dependent, such that ants use different sampling strategies in different situations. Regardless, the presence of a bias suggests lateral specialization within an antennae pair that affects odor sampling on simple trails. This fits with evidence from a variety of studies which has shown a lateralization of insect brains and behaviors (Buchanan et al., 2014; Wes and Bargmann, 2001; Letzkus et al., 2006; Rogers et al., 2013; Frasnelli et al., 2012).

Qualitative and quantitative studies have repeatedly shown that single antennectomy results in a modest decrease in odor gradient- and trail-tracking accuracy in insects (Martin, 1965; Hangartner, 1967; Borst and Heisenberg, 1982; Duistermars et al., 2009; Steck et al., 2010; cf. Lockey and Willis, 2015). Interestingly, we saw altered antennae tip sampling in ants after surgery (Fig. 7A,B). After 2–72 h, the single antenna moved over a larger area of space, including extending across the ant's midline and the contralateral edge of the trail, which is rarely seen in controls (Fig. 7A,C). This is consistent with evidence from Drosophila that show adaptation within days of amputation (Wosnitza et al., 2013; Isakov et al., 2016; Muijres et al., 2017), as well as recent studies on ants showing robust performance across a range of olfactory behaviors after unilateral antenna amputation (Waxman et al., 2017). We found that deviation from the trail (as measured by RMSE) was higher for single-antenna ants (Fig. 7D,F). Ants were, however, still able to track trails along their entire length (15 cm), as has been seen previously, and which is thought to be due to angle corrections after successful temporal comparisons (klinotaxis). Ants with a single antenna made fewer small instantaneous body angle changes and more large turns (Fig. 7E). This is consistent with the idea that single-antenna ants are worse at making small adjustments to stay centered on the trail (i.e. trail following), and subsequently they make large corrections after wandering far off the trail. Indeed, we saw more off-trail and less trail-following behavior in single-antenna ants compared with controls (Fig. S3C). Overall, behavioral changes in antennae movement may partially explain how the single-antenna ant is able to maintain a level of estimated odor information similar to what one antenna of a control ant receives despite greater deviation from the trail (Fig. 7G).

Here, we have developed a better understanding of the multiple behavioral strategies ants use to track odor trails. These include three distinct and stereotyped patterns of body and antenna movement and usage, which affect the amount of odor information received over time by each antenna and the ant brain. These descriptions provide insight into the general strategies ants use for trail tracking as well as evidence for how the two distinct antennae move with respect to each other and relate to motor output. In future studies, behavioral analysis could be coupled to neural recordings to determine how odor inputs and proprioceptive information are transformed into motor outputs in simple, as well as more complex, odor trails.

We thank all the members of the Murthy lab for their helpful discussions on experimental design and data analysis. We thank the members of the Traniello lab, especially Mario Muscedere and Fran Kamhi, for teaching us how to collect, house and dissect ants.