Attack behaviour in naive gyrfalcons is modelled by the same guidance law as in peregrine falcons, but at a lower guidance gain

Abstract

The aerial hunting behaviours of birds are strongly influenced by flight morphology and ecology, but little is known of how this relates to the behavioural algorithms guiding flight. Here we use GPS loggers to record the attack trajectories of captive-bred Gyrfalcons (Falco rusticolus) during their maiden flights against robotic aerial targets, which we compare to existing flight data from Peregrines (Falco peregrinus). The attack trajectories of both species are well modelled by a proportional navigation (PN) guidance law, which commands turning in proportion to the angular rate of the line-of-sight to target, at a guidance gain. However, naïve Gyrfalcons operate at significantly lower values of N than Peregrines, producing slower turning and a longer path to intercept. Gyrfalcons are less manoeuvrable than Peregrines, but physical constraint is insufficient to explain the lower values of N we found, which may reflect either the inexperience of the individual birds or ecological adaptation at the species level. For example, low values of N promote the tail-chasing behaviour that is typical of wild Gyrfalcons and which apparently serves to tire their prey in a prolonged high-speed pursuit. Likewise, during close pursuit of typical fast evasive prey, PN will be less prone to being thrown off by erratic target manoeuvres at low guidance gain. The fact that low-gain PN successfully models the maiden attack flights of Gyrfalcons suggests that this behavioural algorithm is embedded in a guidance pathway ancestral to the clade containing Gyrfalcons and Peregrines, though perhaps with much deeper evolutionary origins.

Steering clear of the confusion effect: aerial predators target fixed points in dense prey aggregations

“…We therefore find no evidence that the hawks engaged in closed-loop pursuit of the bat that they grabbed. As a check on the robustness of this conclusion, we applied an analogous model selection approach to published data on Peregrines and Gyrfalcons pursuing singleton targets 22,23 , which confirmed as expected that their attack trajectories were better modelled by delay-free PN targeting the instantaneous rather than final position of the target (Table S2). As this is the opposite of what we observed for hawks attacking swarming targets here, it is reasonable to test whether the hawks’ observed turns are instead consistent with the use of some form of open-loop steering behaviour.…”

“…A bat is an agile target that cannot be picked out easily within a swarm. However, whereas the question of how predators guide their attacks on singleton targets has been the subject of several algorithmic studies [21][22][23]26 , the question of how they guide their attacks on agile swarming targets has not been addressed previously. Applying the same algorithmic approach to natural behaviour 20 , we find no evidence that hawks attacking swarming bats use closed-loop pursuit of an individual bat, in contrast to raptors attacking singleton targets.…”

“…Fitting guidance laws to empirical data has provided insight into the behavioural algorithms that different predators use to catch lone prey 21-23 . For example, the terminal attack trajectories of Peregrines Falco peregrinus and Gyrfalcons F. rusticolus are well modelled by a guidance law called proportional navigation (PN) 22,23 . This commands the attacker’s turn rate (i.e.…”

“…The quantitative mapping from the relative motion of the target to the steering response of an attacker is called a guidance law 20 . Fitting guidance laws to empirical data has provided insight into the behavioural algorithms that different predators use to catch lone prey [21][22][23] . For example, the terminal attack trajectories of Peregrines Falco peregrinus and Gyrfalcons F. rusticolus are well modelled by a guidance law called proportional navigation (PN) 22,23 .…”

“…We ran these simulations beginning from different points in the flight, from 1.0 s up to a maximum of 20.0 s before the grab, in 0.2 s intervals. We identified the longest simulation for which the relative error was 𝜀 ≤ 0.012, this being the same error tolerance used in previous studies 22,23 . We found that we could model 20 of the 𝑛 = 26 long-range approaches under delay-free PN at 𝑁 = 2 to within this 1.2% error tolerance, with a median path length of 24.5 m (Q1, Q3: 11.9, 45.6 m) and a median duration of 2.1 s (Q1, Q3: 1.4, 3.9 s).…”