Insects that predate aerially usually contrast prey against the sky and attack upwards. However, killer flies (Coenosia attenuata) can attack prey flying below them, performing what we term 'aerial dives'. During these dives, killer flies accelerate up to 36 m/s2. Although the trajectories of the killer fly's dives appear highly variable, proportional navigation explains them, as long as the model has the lateral acceleration limit of a real killer fly. The trajectory's steepness is explained by the initial geometry of engagement; steep attacks result from the killer fly taking off when the target is approaching the predator. Under such circumstances, the killer fly dives almost vertically towards the target, and gravity significantly increases its acceleration. Although killer flies usually time their take-off to minimise flight duration, during aerial dives killer flies cannot reach the lateral accelerations necessary to match the increase in speed caused by gravity. Since a close miss still leads the predator closer to the target, and might even slow the prey down, there may not be a selective pressure for killer flies to account for gravity during aerial dives.

Female killer flies (Coenosia attenuata) were taken from a laboratory colony, established from animals collected in Almeria (Spain). The colony was kept at 60-70% humidity and a 20-25°C temperature in a 12/12 hour light/dark cycle, and was fed live fruit flies (Drosophila melanogaster).

Killer flies were individually placed in transparent vials with only wet filter paper provided for two to three days before testing, to ensure the animals were motivated to hunt. After isolation, two flies were tested in the arena at a time, at room temperature of 20-22°C. The arena was a transparent 160 x 160 x 300 mm box made of 4 mm wide acrylic boards (Perspex^®, Mitsubishi Chemical Lucite Group Ltd, Tokyo, Japan), placed with the long axis resting on a white table. The arena was illuminated with two artificial light sources (4Long-Studio, VelvetLight, Barcelona, Spain) each producing an output of 14,000 lx at 1m, comparable to peak light levels of a moderately cloudy day.

After placing the flies in the arena, a target was moved along the longitudinal dimension of the box, perpendicular to gravity, in either direction. The target was a black 2.1 mm bead, moved between 0.65 and 0.8 m/s on a transparent 0.15 mm thick fluorocarbon fishing line (Vanish, Berkley Fishing, IA, U.S.A.). The line ran along pulleys fixed at the corners of a transparent acrylic support and was moved by a 23HS-108 MK.2 stepper motor controlled through computer software by a ST5-Q-NN DC motor controller (Applied Motion Products, Watsonville, CA, U.S.A.). The target always started moving outside the box. Flies were tested on repeated trials until they stopped responding to the target.

We used two time-synced SA2 Photron cameras (Photron Ltd., Tokyo, Japan) for filming. The recording frame rate used to digitise trajectories was 1000 frames/second, whilst the frame rate to extract wingbeat frequency was 2500 frames/second. The system was calibrated using an altered version of the J.Y. Boguet's Laboratory's MATLAB toolbox (Caltech, http://www.vision.caltech.edu/bouguetj/calib_doc/), running on MATLAB R2014a (version 8.3, MathWorks Inc., Natick, MA, U.S.A.). The movements of the killer flies and target were digitised offline using custom MATLAB scripts for supervised automatic tracking, and smoothed through a fitting algorithm combining trajectory generation using linear fitting with jerk minimisation, run on MATLAB R2012a (version 7.14).

The three dimensional coordinates of targets and killer flies have been exported in CSV files. Each file has three columns, one for each dimension (xyz), and many rows indicating the coordinates over time (intervals of 1ms). Each file name indicates:

• What surface the killer fly took off from to attack (ceiling/floor/wall)
• The number of the attack (attack1, attack2, ...)
• Whether the coordinates are of the target or of the killer fly

The "Calculation.m" file, dowloadable under "Software", contains our dynamical analysis of the attacks to calculate flight power, as illustrated in the article, written in MATLAB_R2020a.