Hysteresis stabilizes dynamic control of self-assembled army ant constructions

Field experiments: collective structures

We found that self-assembled Eciton hamatum bridges adaptively adjust in response to shifts in the terrain on which they are built. Detailed methods are included in Methods: Field experiments. Briefly, we moved foraging trails onto an apparatus where we could introduce a terrain gap. We repeatedly changed the size of this gap by first incrementally increasing it to 30 mm, by 1 mm every 30 s, and then incrementally contracting it at the same rate (See Fig. 1, Methods: Field experiments, and Supplementary Movie 1). As the size of the gap was expanded (the period before the dotted line in Fig. 2a, b) both the volume and number of ants increased to mean maximum values of 1080 mm³ (standard error, s.e. 84) and 18.9 ants (s.e. 1.6), respectively. Ants typically began forming a bridge when the gap was ~5 mm. As the gap size was decreased (period after dotted line), volume and the number of ants decreased back to zero as ants left the bridge. These broad dynamics across the ten complete trials were similar (Fig. 2a, b, inset panels and Supplementary Figs. 2, 3). Additionally, bridge volume (Fig. 2a) strongly correlated with the number of ants in the bridge (Fig. 2b), indicating that the density of ants per unit volume in these structures is relatively consistent (Pearson correlation coefficients range from 0.88 to 0.98 across the ten trials, see also Supplementary Fig. 4). Bridges broke and quickly reformed in eight of the ten trials; breaks occurred in both experimental phases, and these broken periods were excluded from analyses. Overall, these results show that bridges adjust dynamically to changing terrain geometry, as stretching the bridges caused them to become larger, with more ants, and contracting bridges caused them to become smaller, with fewer ants.

However, these changes were not symmetric—adjustments in the contraction phase were not the inverse of adjustments in the expansion phase. We found consistent hysteresis in several metrics; for a given gap size, bridges were larger and made up of more individuals during the contraction of the gap than the expansion (Fig. 2c, d; t-test for volume: mean extent of hysteresis = 0.43, 95% CI = 0.29 to 0.58, t = 6.7, df = 9, p < 0.0001; t-test for number of ants: mean extent of hysteresis = 0.28, 95% CI = 0.13 to 0.43, t = 4.2, df = 9, p = 0.002; see Methods: Data Analyses). This “excess” mass went into an increased mean cross-sectional area in contraction-phase bridges (Fig. 2e; t-test: mean extent of hysteresis = 0.40, 95% CI = 0.26 to 0.53, t = 6.7, df = 9, p < 0.0001) and caused contraction-phase bridges to hang lower, having more slack, than expansion bridges (Fig. 2f; t-test: mean extent of hysteresis = −0.27, 95% CI = −0.34 to −0.21 t = −9.1, df = 9, p < 0.0001). Bridges had the most slack midway during the contraction (Fig. 2f), suggesting that as the gap begins closing the bridge does not immediately shorten, instead of hanging lower until shortening later in the contraction phase. Hysteresis in each of these metrics was present in every trial, regardless of variance in traffic or the presence and timing of breaks.

This hysteresis cannot be explained by traffic, which is known to affect bridge structures²⁴, as traffic was independent of phase (Supplementary Fig. 9). One may speculate that hysteresis emerged because we changed the gap size too rapidly for the bridge to keep up. Yet this is also an insufficient explanation, as earlier work demonstrated that ants in bridges adjust to changes in traffic flow within seconds²⁴. We also excluded the possibility that delayed response to geometric changes causes hysteresis, as discussed in our mechanistic models section, below. Thus, the state of these structures depends not only on gap size and traffic, which was already known but also on the history of the structure, which has strong implications for our understanding of the control mechanism. Additional measures of the bridges, as well as full-size plots for each trial, are included in the supplementary information (Supplementary Figs. 2–8).

Hysteresis is a type of structural memory²⁷, because the current state (i.e., number of ants) depends not only on the value of a driver (i.e., gap size) but also on the history of the system. Hysteresis is an important feature of many biological systems, including cellular and genetic systems^28,29, neurological systems^28,30, animal groups^31,32,33,34, and ecosystems^35,36. While hysteresis in some contexts is a signal of reduced resilience (e.g., ecosystems³⁷), it is a key stabilizing feature of other systems, especially systems designed or evolved for control. Hysteresis can allow for multiple stable states (bistability)^33,38,39 and memory^39,40 and can reduce the impact of noise⁴¹. Biological examples include cellular signaling—where both bistability and memory are associated with hysteresis^38,39,40,42—and human vision—low corneal hysteresis seems to exacerbate glaucoma⁴³. Hysteresis is also intentionally added to stabilize many engineered control systems, for example, to make these systems immune to noise, preventing constant on-off switching near thresholds (e.g., thermostats, Schmitt triggers⁴⁴). In army ant bridges, hysteresis may play a similar role, improving stability and allowing the structures to avoid unnecessary adjustments to small, momentary gap size changes. We investigated this possibility as described in the following section.

Mechanistic models

Living army ant bridges adaptively respond to changes in the geometry of the surrounding terrain. To understand how this response is dynamically controlled, we developed three behavioral models of individual decisions to join or leave structures—two linear with respect to the gap perturbation and one nonlinear—as mechanistic hypotheses. Note that joining and leaving events do not appear to be directly prompted by changes in gap size, as the density of events is uniform across the 30-s intervals between adjustments (Supplementary Fig. 10). We focused first on whether simulations of these models under the conditions of our field experiments can reproduce the hysteresis we observed, which cannot be simply explained by factors previously reported to affect these structures, such as traffic and gap size. We then simulated these models’ behavior when built on terrain that shifts slightly back and forth, as we expect these vibration-like movements to be common in the ants’ environment and hysteresis is a known mechanism for damping oscillations. We were particularly interested in whether nonlinear mechanisms are required to explain the collective bridge dynamics because nonlinear responses are an important feature of many types of complex emergent phenomena^34,45,46 and implicated in engineered hysteresis^47,48,49,50. All three models were parameterized using the data; none of the models include any free parameters that we tuned.

Linear models: baseline and delay

We explored two data-driven linear models of ants joining and leaving bridges (Fig. 3). These are fixed-probability models because they rely on the assumption that individuals’ probability of joining and leaving is constant during each phase. It is reasonable to expect such constant behavior, as the rate of perturbation in our experiments was constant at 2 mm/min. We simulated the number of ants in a bridge resulting from stochastic joining and leaving events, with the restriction that the number of ants leaving cannot be larger than the current number in the bridge. Each simulation timestep corresponds to a 30-s gap interval in the real experiments. The numbers of joining and leaving events in each timestep were drawn from negative binomial distributions fitted to the data (see Fig. 3a and Methods: Models). During the expansion, the mean number of joining and leaving events in each 30 s timestep were 1.13 and 0.67 (dispersion parameter, θ = 1.33 and 1.98), respectively, and during the contraction, there were an average of 0.33 and 1.04 joining and leaving events (θ = 0.65 and 8.10) in each step, respectively.

Our linear baseline model assumes that ants join and leave bridges quickly after deciding to do so, while in our linear delay model there is a lag of constant duration between an ant deciding to leave the bridge and actually doing so. As discussed above, long delays may be unlikely based on known response times to traffic changes²⁴, yet we include this possibility because one may intuitively expect it to result in hysteresis. We modeled leaving delays of 30 s up to 3 min (1 to 6 timesteps), and we do not allow for a delay in joining, as ants that could potentially join the bridge (those walking over it), do not remain nearby for long enough to result in a joining delay of more than a couple of seconds. If there were indeed a leaving delay in real ants, then the leaving events recorded at the beginning of the contraction phase must reflect decisions to leave made during the last part of the expansion. Thus, we adjusted the distribution for leaving events in our delay model to still accurately reflect leaving decisions for the assumed length of delay (see Methods: Models). We evaluated both of these linear models by first plotting the average simulated expansion and contraction dynamics to compare with the experiments, and then by measuring the extent of hysteresis. The extent of hysteresis can range from −1 to 1; we define it as the area between expansion and contraction curves as in Fig. 2d, scaled by the area under the higher curve, and it is negative if the expansion curve is higher.

The baseline model failed to reproduce the dynamics we observed in real ant bridges. While the models’ behavior was highly variable, the majority of simulations resulted in negative hysteresis, that is hysteresis in the opposite direction as the data, with more ants in the simulated bridges during the expansion than the contraction (Fig. 4a). The mean extent of hysteresis was −0.22, and it was negative in 65% of the 10,000 simulations, compared with an average of 0.28 in the real experiments, with 100% of trials having positive hysteresis (distribution of extents of hysteresis shown in Supplementary Fig. 11). The negative hysteresis produced by the linear models is due to a slight asymmetry in the data on joining and leaving in each phase. Briefly, in our experiments, the net loss of ants (number leaving the bridge minus number joining) in the contraction (specifically, 0.71) tended to be slightly higher than the net gain during the expansion (0.46, Supplementary Fig. 12). The delay model performed only slightly better than the baseline model, resulting in an average extent of hysteresis of −0.15, and the hysteresis was negative in 59% of simulations (Fig. 4b and Supplementary Fig. 11). The delay model performed poorly regardless of the length of delay, and the results of models with delays of one step up to six steps were all similar (Supplementary Fig. 13).

We further tested these models by subjecting them to small, momentary changes in gap size. We expect that these vibration-like conditions occur frequently in nature, as bridge anchors such as leaves shift slightly in the breeze. We performed 1000 simulations of experiments with three phases of movement: expansion as before, a vibration of 1 mm (gap shifts back and forth from 29 and 30 mm), and contraction as before. The behavior of both models is highly variable, with many of the simulated bridges rapidly growing or shrinking to unreasonable numbers of ants during vibration. Even though the gap size was always 29 or 30 mm throughout the vibration phase, the baseline and delay models resulted in “extreme” bridges, at some point having fewer than five ants or more than 25 ants, in 84 and 85% of simulations, respectively. Both models also tend to slowly lose ants on average during vibration (thick black line in Fig. 4d, e), which would lead to bridge failure over time; this is again due to the slight asymmetry in joining/leaving between phases, discussed above (Supplementary Fig. 12). Furthermore, in 41 (baseline) and 39% (delay) of simulations the structures reached zero ants, indicating that with the linear models, even a slight asymmetry leads to frequent failure during vibration. These models clearly do not produce structures that could reliably support traffic when facing the small, repeated shifts in gap size which are likely common in nature.

Given the uniform rate of perturbation in the field experiments, it was reasonable to expect ants to have fixed probabilities of joining or leaving in each phase. However, both linear models failed to reproduce hysteresis resembling our experimental results. These models propagate errors, in the sense that if too many ants leave in one timestep, leaving events are still equally likely in the next. Modeling a delay in leaving does not resolve this; every gap shift gets a response whether or not it should, even if that response is delayed. The linear models are also highly sensitive to asymmetry—even the slight asymmetry in the net gain/loss of ants between phases in our experiments led to these models typically producing negative hysteresis and bridges which shrank during vibrations. Finally, both linear models resulted in such variable behavior—even with small, momentary gap shifts—that any given simulation was unlikely to produce a viable structure.

Nonlinear accumulator model

After excluding the linear models, we chose to explore a nonlinear “accumulator” model, with parameters still drawn from the data. This is a dynamic-probability model—the chances of joining and leaving are based on the excess or deficit in the number of ants in the bridge, that is, the deviation of the actual number from the equilibrium number for that gap size (calculated using equilibrium “packing density”, or ants per mm of gap, as described below, also see Methods: Models). When the number of ants is small for the gap size (i.e., the bridge is stretched and packing density of ants is low—large deficit), ants are very likely to join a bridge and unlikely to leave. Conversely, when there are more ants in the bridge than necessary (i.e., the bridge is slack with excess ants), ants are likely to leave and unlikely to join. We refer to this as an “accumulator” model because as changes in the terrain geometry accumulate, the likelihood of response increases nonlinearly as described below, and small changes are relatively unlikely to produce a response. While there are endless possible nonlinear models, we chose this model to explore because it agrees with our intuition that bridge adjustments respond to the magnitude of perturbations. It also fits earlier work showing that the packing density of bridges affects individual behavior²⁴, and conceptually related models explain other social insect behaviors, such as response threshold⁵¹, error correction²³, and integrator⁵² models.

We parameterized our accumulator model using our experimental data. The equilibrium packing density of bridges in our experiments was 0.51 ants per mm of gap size; this is the packing density at which ants are equally likely to join as leave a bridge (Fig. 3b). Thus, a bridge spanning a 20 mm gap should be composed of about ten ants, and if our simulations result in a 20 mm bridge with only seven ants, we model a deficit of 3. We then mapped simulated differences from this equilibrium to the resulting number of joining and leaving events using our experimental data. We used negative binomial regressions of the data to predict the mean number of each type of event, µ, at each deficit value—thus, µ depends nonlinearly on the deficit/excess (∆(t)), as shown in Fig. 3c. Specifically, joining events increase with increasing deficit (β = 0.049, θ = 0.773), and leaving events decrease (β = −0.067, θ = 1.96); see Methods: Models for analysis details. These data also confirm that high deficits lead to more ants joining and fewer ants leaving while the inverse is true for bridges with excess ants, as expected for the accumulator model. We also note the data include more instances of extreme excess than instances of extreme deficit—while it was somewhat rare in our experiments for a bridge to have a deficit of more than five ants, it was not uncommon to have an excess of ten or more ants. This is likely because if the deficit grows too large a bridge will fail. On the other hand, while a very large excess (a highly negative deficit) may be inefficient, it is unlikely to lead to bridge failure.

The accumulator model behaves similarly to the experiments with real ants (Fig. 4c). The average extent of hysteresis was 0.48 (see Supplementary Fig. 11 for distribution of hysteresis extents). While this is higher than the extent of hysteresis in the experiments with real ants, which was 0.28, it is much closer than the linear baseline and delay models. Furthermore, 96% of the accumulator model simulations resulted in positive hysteresis, compared with 100% in the real experiments. The accumulator model also performed well in the vibration condition (gap size-shifting back and forth between 29 and 30 mm). Unlike the linear models, this model maintained a consistent number of ants on average during vibrations (thick black line in Fig. 4f), and with far less variation—75% of simulations avoided “extreme” bridges, maintaining between 5 and 25 ants for the entire vibration phase, and only 1% of simulations reached 0 ants at any point. This is because the 1 mm shifts in the vibration phase hardly change the packing density, so joining and leaving events are relatively unlikely; this is quite different from the linear models which are equally likely to respond to every 1 mm shift. Furthermore, in contrast to the linear models, in which the chance of joining and leaving events is unaffected by past behavior, the accumulator model is self-correcting—if too many ants leave in one timestep the deficit increases, making future leaving events less likely and joining events more likely. Thus, this model’s behavior inherently depends on its history, and this hysteresis damps potential oscillations near the equilibrium, making the bridges more stable.

Finally, we explored how sensitive the accumulator model’s behavior is to the maximum size of the gap and to the rate of gap size change. We found that this model is scalable, as it behaves similarly whether the bridge spans a gap reaching a maximum of 3, 6, or 9 cm (Fig. 4g). The behavior of the model is somewhat sensitive to the rate of gap size change. Specifically, the extent of hysteresis is rate-dependent: when gap changes accumulate more quickly there is predicted to be more relative hysteresis (Fig. 4h), and we also predict slightly smaller bridges in this case. Yet the overall system dynamics are similar across rates.

The dynamic-probability accumulator model outperformed the fixed-probability linear models in every analysis. It consistently reproduced hysteresis in the same direction as our experiments, performing well with the relatively large, directional changes we exerted in our real experiments. It maintained stable bridges when facing small, momentary changes in gap size, avoiding constant adjustments, because it is self-correcting, which also makes it much less sensitive to asymmetry. Furthermore, it is scalable, behaving consistently across different maximum gap sizes, and not highly sensitive to the rate of gap size change. Whereas the poorly-performing linear models depend only on external dynamics (i.e., expansion vs contraction), the accumulator model incorporates nonlinear feedback that produces stabilizing hysteresis. Nonlinear dynamics also play a key role in many other mechanisms of emergent action and control, due to quorums, thresholds, critical points, and more^34,45,46, and the accumulator model is similar to, though mechanistically distinct from, an integrator model previously applied to ant foraging⁵². We conclude that this nonlinearity is an important feature of the dynamic control mechanism of collective army ant bridges, which are able to withstand small perturbations in gap size without adjusting, becoming more likely to adjust as perturbations accumulate.

Field experiments: individual cues

According to our model, the deviation from equilibrium packing density of ants in a bridge (i.e., magnitude of deficit or excess) can explain the number of ants joining and leaving these structures, but we do not expect ants to respond to this directly, as it may be difficult for ants to perceive. Here, we consider immediate cues causing individual ants to respond to such a deficit—how can an ant locally detect if a bridge is overstretched or slack? We investigate performance (i.e., how easy the bridge is to cross) as a likely candidate cue for ants sensing the deviation from equilibrium. We hypothesize that bridges with a large deficit (too few ants in the bridge for the gap size) would be harder to cross and have low performance. We further assume ants joining a bridge that has a large deficit will improve its performance (i.e., whether and how easily ants can walk across it), and ants leaving a slack bridge, with excess ants, will likely also make it easier and/or faster to cross. Thus, while there are many possible cues that may lead to individual joining and leaving decisions, we first consider performance-based cues. Intuitively, both joiners and leavers may have information about bridge performance; potential joiners (those walking on the trail) can likely sense how easy the bridge is to cross while potential leavers (those within the bridge) may be able to sense how many and how rapidly ants are able to cross over them²⁴.

Our measure of bridge performance is the flow rate of ant traffic over the bridge itself relative to what would be expected based on the current flow rate on the trail next to the bridge (while the total flow of ants must be consistent across the trail, local flow rate changes with fluctuating traffic). We used optical flow as a proxy for the local flow rate, which integrates information about the speed and number of crossing ants, and we account for overall changes in traffic by recording performance as the signed residuals from regressions of optical flow on bridges vs optical flow on both of the adjacent platforms (Fig. 5, Supplementary Note 2, and Supplementary Fig. 14). Our measure of performance reflects the ease with which ants can cross a bridge. If a bridge slows down the flow of ants to a greater extent than usual, the performance measure will be low; high performance reflects bridges which do not slow down traffic or do so less than usual. In general, performance was higher in the expansion phase than the contraction phase (reflected in the distributions in Fig. 6). This suggests that the taut bridges occurring during expansions are easier to cross than more slack bridges in the contraction phase. We investigated whether performance is a cue for joining or leaving events by comparing performance immediately before events with performance at random intervals, using a resampling approach (see Methods: Data analyses).

Joining mechanisms

We found that bridge performance strongly affects ants’ decisions to join, yet the direction of this effect is contrary to our expectations. We predicted low performance to occur at high deficits (fewer ants than the equilibrium number) and then lead to joining, as ants struggling to cross a bridge will slow down, and will likely then be walked on by those behind them. However, performance immediately preceding joining events in our experiments was substantially and significantly higher than would be expected by chance (Fig. 6, left). This result cannot be attributed to overall traffic patterns, which were previously known to affect individual decisions^24,25, because our performance measure accounts for overall traffic. The effect of high performance on joining is distinct from—and in our experiments, clearer than—the overall effect of traffic (Supplementary Fig. 15).

The evidence that joining events are preceded by high bridge performance initially seems to contradict our data- and model-based results that bridges with high deficits have higher joining rates. This discrepancy arises from our faulty assumption that bridges with high deficits (i.e., too few ants) will have lower performance. In fact, high deficit was not associated with low performance in our experiments (Supplementary Fig. 16), and we saw higher performance on average during the expansion than the contraction phase, suggesting that having excess ants in a bridge may be worse than having too few, at least within the range of packing densities we observed. While we expected the equilibrium packing density to be close to optimal, it seems to result in more ants than necessary for high bridge performance. Additionally, earlier work indicates that after natural bridges are broken, traffic flow reaches the previous level before the number of ants in the bridge recovers to the previous number²⁴. Ants continue to join these bridges that are back to “normal” traffic flow, further supporting that ants will join bridges with high performance. Our results also strongly support that ants join bridges with high performance. This positive feedback may reflect colonies’ investment in portions of the trail that are functioning well, regardless of the bridge’s deviation from equilibrium. After all, ants should be capable of abandoning path sections that are not functioning to find new paths. These “extra” ants may also strengthen the bridges to minimize breaks, may provide extra capacity for sudden traffic increases²⁴, and may provide some insurance against the gap increasing.

Leaving mechanisms

We do not observe strong effects of performance on decisions to leave bridges. While performance immediately prior to leaving events is significantly lower than expected during the expansion phase (Fig. 6, right), the effect is much weaker than for joining events and is sensitive to the time interval over which performance is measured (Supplementary Fig. 14). There is no such effect of performance on leaving events in the contraction phase. Thus, potential leavers do not seem to use performance as a cue.

Instead, we hypothesize that potential leavers respond more directly to a bridge’s deviation from the equilibrium number of ants, by sensing a closely-linked structural feature such as how taut or slack the bridge is. Intuitively, we expect ants in a bridge to be more stretched and experience higher tension at high deficits (i.e., low packing density) than those in a bridge with excess ants. We explored this hypothesis at the level of the entire structure and found that the tautness (height of the center of mass) of an entire bridge does not predict leaving events in either the expansion or contraction phases (Supplementary Fig. 17).

However, this global analysis is insufficient, because structural or mechanical cues experienced by an ant will vary depending on where the ant is located within a bridge and the network of grasping connections in the structure (which is known to affect the probability of leaving²⁴). We, therefore, expect considerable variation in the tension (or other mechanical factors) experienced by individuals within a structure; for example, even in a slack bridge ants at the edge acting as anchors may experience high tension. We leave a direct test of the hypothesis that structural cues prompt leaving events to future studies that allow for careful modeling of the forces on individual ants. Nevertheless, this hypothesis has some indirect support. Consider that in addition to ants experiencing lower tension on average when a bridge is slack, there should also be a reduction in tension when the bridge is supporting less mass, i.e., when traffic on the bridge is lower. We know from earlier work that ants are more likely to leave when they are walked on less²⁴, and this is further supported by our experiments (leaving events associated with low absolute bridge traffic, Supplementary Fig. 18). If potential leavers use reduced tension as a cue it may account for both the known effect of traffic on leaving and our results that ants are more likely to leave when there are excess ants in the bridge.

This points to a clearer explanation of our results, with two separate mechanisms accounting for joining and leaving (Fig. 7). As discussed above, we know that E. hamatum workers join bridges with high relative traffic flow (high performance), suggesting that colonies may invest in high-functioning sections of trail. This does not contradict our model, as plenty such high-performance periods will occur when the deficit is also high; large deficits do not lead to low performance as we assumed. As ants join a bridge the deficit is reduced, and if excess ants join it may prompt those ants or others to leave, perhaps because they experience low stretch or tension due to the extra ants. Joining and leaving mechanisms do not both need to be calibrated such that bridges with too many ants get smaller and those with too few get larger, if the combined effect of joining and leaving produces this response. Thus, by potential joiners and leavers responding to different cues, these dynamic bridges can both adjust in size to reflect traffic requirements and respond to trail dynamics at a larger scale, allowing investment in path options that perform well. Army ants also use distinct sets of cues for joining and leaving defensive walls that form to separate trails from competitors²².

While these results explain how well-functioning bridges respond to shifts in terrain geometry, it is less clear from our study how poorly-performing bridges recover from periods of high deficit. Such periods could be rare if trail routes shift so gradually that performance stays high; yet we still intuitively expect them to occur, for example when a bridge is first needed. What causes ants to join bridges with too few ants and low performance? One potential explanation is that there are one or more other joining cues in addition to high performance, such as social cues or the presence of pot-hole-like gaps in the bridge surface¹⁸. By focusing more on bridges as they initially form, including the recovery of bridges from breaks, future experiments could determine whether additional joining cues are important.

Eciton hamatum army ants self-assemble into remarkably flexible, living bridges, capable of adjusting to constant changes in real time, as structures rapidly adapt to terrain disturbances, forming and growing to 3 cm in just 15 min in our experiments. These adjustments allow colonies to maintain robust trails with heavy traffic in unpredictable conditions, and the response mechanisms may extend to other army ant structures, including ladders, ramps, scaffolds, pot-hole plugs, and bivouacs^{15,16,17,18,19,20,23}. At a global level, army ant colonies can behave like living organic materials, exhibiting cooperative behaviors that are analogous to those seen across scales, including in morphogenesis, wound healing, and the assembly of slime mold fruiting bodies^8,53,54. For example, the bridges we studied add and remove material (ants) as needed to maintain the high flow of ants on the trail; in this sense, their assembly resembles epithelial tissue morphogenesis, in which cell division is induced in stretched tissue (adding material), and apoptosis is induced in slack tissue (removing material)^5,6.

We examined how living army ant structures respond to the dynamic control problem of the ground moving beneath them, focusing on whether and how the structures adapt, what the underlying control mechanism is, and how this control is implemented. In nature, these bridges face both long-lasting terrain shifts from sources such as animals moving through the leaf litter as well as frequent small, vibration-like movements as anchors move back and forth in the breeze. We found that the number of ants making up a bridge is controlled collectively by ants moving along the trail—who join high-performing bridges—and by ants making up the bridge—who leave when a bridge has too many ants. These separate mechanisms for joiners and leavers balance one another and lead to hysteresis, an important feature in the control of both biological and engineered systems, which adaptively stabilizes the bridges, preventing constant switching between ants joining and leaving. Our accumulator model reproduces overall patterns similar to real bridges including this hysteresis—which our experiments showed is a defining feature of the system—and is self-correcting; producing robust bridges which stably adjust to lasting terrain shifts while disregarding momentary, vibration-like movements. By illuminating the presence and benefit of stabilizing hysteresis in army ant bridges, our work demonstrates the potential of related hysteretic mechanisms for stable control of decentralized systems—in which noisy sensing, decision-making, and actions must all be coordinated across dispersed individuals. Fundamentally, such decentralized control drives the success of these systems across scales, from subcellular organization to superorganisms.

Source: Ecology - nature.com