Aerodynamic imaging by mosquitoes inspires a surface detector for autonomous flying vehicles

See allHide authors and affiliations

Science  08 May 2020:
Vol. 368, Issue 6491, pp. 634-637
DOI: 10.1126/science.aaz9634

Sensing surfaces like a mosquito

Although sonar or lidar are used by autonomous vehicles to detect nearby objects, these approaches incur significant equipment and signal-processing costs. Nakata et al. show that mosquitos detect surfaces using the flow fields caused by the movement of their own wings (see the Perspective by Young and Garratt). Near surfaces, there are small changes in the pressure and velocity that mosquitos can detect using their sensitive antennae. The authors translated this process into a simple, low-cost approach for detecting surfaces near a flying quadcopter.

Science, this issue p. 634; see also p. 586


Some flying animals use active sensing to perceive and avoid obstacles. Nocturnal mosquitoes exhibit a behavioral response to divert away from surfaces when vision is unavailable, indicating a short-range, mechanosensory collision-avoidance mechanism. We suggest that this behavior is mediated by perceiving modulations of their self-induced airflow patterns as they enter a ground or wall effect. We used computational fluid dynamics simulations of low-altitude and near-wall flights based on in vivo high-speed kinematic measurements to quantify changes in the self-generated pressure and velocity cues at the sensitive mechanosensory antennae. We validated the principle that encoding aerodynamic information can enable collision avoidance by developing a quadcopter with a sensory system inspired by the mosquito. Such low-power sensing systems have major potential for future use in safer rotorcraft control systems.

At night, in caves, or in otherwise visually compromised environments, animal guidance and control systems must sense and avoid obstacles without relying on optical information. Mechanoreceptors in arthropods are extraordinarily sensitive and diverse (1), and insects exploit this fully (2), including for the detection of self-induced flows. For example, fields of unidirectional trichoid sensilla are likely a key component of the fused sensory input used by flying insects to monitor their attitude (3), and changes in forward speed can be regulated by aerodynamic drag on the antennae (4). In insects, antennal motion is detected by the Johnston’s organ (JO), an array of chordotonal mechanoreceptors located in the antennal pedicel. The JO can detect fluid flows, gravitational pull, and acoustic stimulation, and it is one of the most sensitive mechanoreceptive organs in the animal kingdom (5). Mosquitoes have exceedingly sensitive JOs. The radial organization of their ~12,000 mechanoreceptive units functionally arranged in antiphase pairs (6) allows mosquitoes to respond to antennal deflections of ±0.0005° induced by ±11-nm air particle displacements in the acoustic near field (Toxorhynchites brevipalpis) (7) or to acoustic particle velocities of ~10−7 ms−1 (Culex quinquefasciatus) (8).

We took inspiration from such neurophysiological evidence and postulated a sensory mechanism for C. quinquefasciatus that can explain recent behavioral experiments showing that mosquitoes avoid surfaces invisible to their compound eyes (9). The absence of visual cues indicates that another source of close-range information exists, and we hypothesized that these alternative cues are manifest within interactions between the fluid and antennae or hair structures. Specifically, we propose that mosquitoes can detect changes to their self-induced flow patterns caused by the proximal physical environment. These changes to the downwash flow patterns initially generated by the flapping wings arise as the jets of air impinge on the obstacle’s surface. This noncontact, sensory modality for flying insects is somewhat akin to the hydrodynamic imaging capability of the lateral line system in fish (10, 11), which is also fundamentally a fluid-dynamic, pressure-based system. Such a system would be particularly useful for mosquitoes, which must be adept at stealthy landings on hosts (12) and depositing eggs over water at night.

We demonstrate how nearby surfaces may be detected by mosquitoes by means of the flow field produced during flapping flight (13), which is modulated in response to surfaces at magnitudes sufficient for detection by their mechanosensors. We implemented these governing principles in developing a miniature flying vehicle that operates close to the ground and walls and is fitted with a sensor package that can detect surfaces at distances sufficiently far from collision for effective obstacle avoidance (movie S1).

Mosquito wingbeat kinematics show high wingbeat frequency (wbf), low wingbeat amplitude, and large, rapid spanwise rotations. These features result in unorthodox aerodynamic flows around the wings themselves (13) and two concentrated jets of fast-moving air that merge approximately two wing lengths beneath the body. Because of the shallow stroke amplitude, the jets are more focused than the wake of other flying animals; this may help to improve the signal if the interaction of the induced flow with a ground plane is important for collision avoidance. Building on our previous dataset (13), we performed further computational fluid dynamics (CFD) simulations at a range of distances from either the ground or a wall plane to quantify the effect on local flows around the mosquito (Fig. 1A and fig. S1). Movie S1 shows flow simulations at infinite altitude (where “infinite” in this case means flight at an altitude far from a surface) and when the jets impinge on a ground plane 10 mm below the mosquito.

Fig. 1 Velocity and pressure distributions around mosquitoes flying near surfaces.

(A) Front view of a mosquito hovering at five altitudes measured from the mosquito body with downwash shown in blue and upwash in red. The flow visualization plane is shown at maximum wingspan. A discrete jet from each wing merges in the infinite and high-altitude cases. (B and C) Side view of a hovering mosquito (gray) and distribution of absolute wingbeat-averaged mean difference in pressure relative to the infinite case |ΔP|¯ (Pa) measured in the sagittal plane. The pressure distribution in free airspace is compared with flight near a wall [(B), where the wall is the left edge of the panel] and at varying altitudes (C); white crosses show monitoring location corresponding to the tip of the antenna. (D) Particle velocity detection threshold of the male JO showing a secondary notch of enhanced sensitivity (white arrow) within the male wbf range [see the supplementary materials for electrophysiology methods; also see (8)]. Gray shading indicates the range of male wbf’s observed during free flight. The JO’s secondary notch has a particle velocity sensitivity shown by the solid line. The primary notch at ~200 Hz is used for mating communication and is tuned to tones generated by the male-female wbf distortion product. (E) The amplitude of change in velocity magnitude at wbf measured at the antennae increases with proximity to the ground. A straight line of best fit is plotted (blue, with 95% confidence intervals shown as dashed lines) to show the intersection with the JO flow velocity sensitivity at the male wbf alone (solid horizontal line). (F) Amplitude of changes in velocity magnitude at the antennae in the frequency domain, calculated as the fast Fourier transform (FFT) at infinite altitude subtracted from the FFT at a given altitude over 50 wingbeat cycles. Differences are always greatest at wbf irrespective of altitude. Asterisk indicates JO particle velocity sensitivity at wbf.

Downwash dominates the flow field at higher altitudes. However, at lower altitudes (<10 mm), the downwash velocity progressively decreases and recirculation can be seen in some regions, particularly under the body. To see this effect more clearly, we calculated the wingbeat-averaged pressure deltas for each distance relative to the infinite altitude case (Fig. 1C). The zones with the largest pressure deltas were located below the thorax and, unexpectedly, above the head. The antennae, with their sensitive JO at the base (7, 8), are therefore well placed to measure subtle changes in the vector strength of particle velocity in the anterodorsal region of the head despite being located the farthest from the ground. Flow-sensitive hairs along the hind leg femur, and elsewhere, could also detect changes in flow velocity associated with these pressure changes, especially at the lower altitudes, although hind leg hair sensitivity is an order of magnitude lower (fig. S2). Mosquitoes extend their hind legs toward a surface when landing and backward when flying, which complements the JO in detecting pressure differences caused by ground and wall effects. The antennae of flying insects are self-stimulated both by periodic air movements caused by wingbeats and by tonic flow from translation through the air. Recent mosquito tuning data show two sensitivity peaks in male JOs. One occurs at lower frequencies (centered at ~280 Hz) and is tuned to detect the wbf of females using an acoustic distortion mechanism (8). A secondary peak of sensitivity is centered on frequencies similar to those at which males fly (600 to 800 Hz), which would enable a male mosquito to hear its own flight and possibly that of other nearby males (8, 14). Male mosquito JOs are therefore adept at perceiving tiny changes in the direction and magnitude of flow velocity of the type associated with proximity to surfaces, potentially using one sensitivity band to detect females and another for detecting changes to their self-generated flow fields when encountering obstacles. In addition to ground effects, wall surfaces also modulate the flow field (Fig. 1B). Again, changes in pressure distribution can be seen above the head and below the thorax, so both floors and walls could be detected by the same cuticular flow sensors or pressure sensors.

At the male wbf, the JO exhibits a local peak in sensitivity and can detect changes in flow velocities on the order of 10−4 ms−1 (Fig. 1D and supplementary materials). We include this empirically derived limit in Fig. 1, D to F, where we present the change in flow velocity at the wbf with varying proximity to the ground (Fig. 1E) and the frequency spectrum of the induced flows (Fig. 1F). Flow velocity oscillates less with altitude, and closer proximity to the ground does not cause oscillations in the flow experienced by the JO to deviate from wbf. At higher altitudes, differences in the magnitude of velocity fluctuations at the wbf become less pronounced and, for numerical reasons, CFD will eventually fail to capture the very smallest changes in velocity. There is a considerable computational burden as the fine mesh extends to ever more distant ground planes and the velocity deltas tend to zero; nevertheless, a clear trend can be seen whereby the JO can easily detect changes at low altitude, but with a diminishing response as the altitude increases, until the threshold for detection is not met (Fig. 1E).

The intercept of the CFD-derived velocity changes and the measured sensitivity of the JO predict a maximum surface detection distance in Culex mosquitoes of 36.4 mm, or 20.2 wing lengths. This is a conservative estimate because it only considers the content of the flow signature at wbf. This distance predicted for Culex males is broadly consistent with egg-laying dipping behavior in female Anopheles, which dip to altitudes of 20 to 70 mm above the water surface (9). Detection of a ground plane at such distances is far in excess of that which might be expected by the ground effect typically referred to in the aerodynamic literature, where notable improvements in lift and drag force characteristics of wings become negligible beyond an altitude of a single wing length or rotor radius. In our mosquitoes, the negative pressure delta region observed above the head and under the thorax when close to the floor occurred as a result of increasing unsteadiness of the flow in this region, leading to higher peak velocities and lower pressures (Fig. 1C). Conversely, away from surfaces, the flow around the body was relatively steady because the speeds of the wing bases were low.

Mosquitoes are not known to have pressure receptors that could monitor the reflected sound from nearby surfaces in the same manner as echolocating animals. Although we do not rule out the possibility that the JO could detect the reflected particle velocity component of self-induced sounds, it would be less useful than the pressure component because the particle velocities decrease with the inverse cube of distance rather than the inverse square. Moreover, the frequency of the flight tone means that the wavelength of the acoustic signature is relatively large, on the order of 0.5 to 1.0 m, which limits precision in locating a surface. By contrast, typical echolocation in gleaning bats uses frequencies in the tens of kilohertz, giving a superior resolution by two orders of magnitude. Given the relatively large changes in particle velocity induced by each wingbeat that can comfortably be detected by the JO at altitudes of many body lengths, we think that this is a more robust solution to surface detection than echolocation.

To show how mechanosensory flow-field monitoring can be used in collision avoidance in autonomous systems, we fitted a small quadcopter platform with a bioinspired sensor that can detect floors and walls using physical principles similar to those described above: specifically, modulation of a deforming flow field. It is lightweight, power-efficient, and stealthy, with no additional emission of light or electromagnetic radiation necessary. It is also applicable to rotorcraft or flappercraft of any scale and can work in conditions that are unsuited to alternative range-finding tools. We instrumented an existing 27 g platform (Crazyflie 2.0, Bitcraze, Sweden) with custom circuits and algorithms to identify obstacle proximity based on pressure sensor readings. The stand-alone sensor module performs reliable obstacle detection up to three rotor diameters away during autonomous flights.

The device, like the mosquito, will be most sensitive if sensors are mounted at locations experiencing the greatest changes in the flow field when approaching surfaces. Nearby surfaces distort the flow field all around the body, making surface detection simple, direct, and robust; however, to determine optimal sensor design, number, and placement, it is necessary to find the most affected regions. We used stereoparticle image velocimetry to measure fluid velocities around the quadcopter at various altitudes and proximities to a wall (Fig. 2 and fig. S3). These flow measurements were used to inform the position of probe tubes relative to the annular jets and regions of recirculation under the control boards. The probes were connected to differential pressure sensors, which are a more accessible solution than particle-velocity probes (Fig. 3 and figs. S4 to S7). Because the dynamic pressure is proportional to the square of flow velocity, the same physical phenomenon underpins the sensing capability. Ground effect could be detected using a pair of probes extending above and below the craft, and the direction of nearby walls could be detected by using paired probes extending fore to aft, laterally, or diagonally. Further detail on the design criteria and the pressure delta thresholds for each proximity condition are detailed in supplementary materials.

Fig. 2 Quadcopter flow-field characterization.

(A) Slices showing induced downwash for a quadcopter hovering at a range of altitudes in multiples of rotor diameter (D = 46 mm). Line integral convolution shows instantaneous streamlines and color flood shows vertical velocity. (B) Differences in velocity magnitude at a range of altitudes. (C) Schematic of the craft showing the particle image velocimetry measurement plane (red) with respect to a center line (dashed). (D) Oblique and (E) Top view of the three-dimensional flow field at an altitude of 2 D. Four annular jets emanate from the rotors and recirculate under the fuselage (isosurface of downwash and upwash: 4 ms−1 in red; –2 ms−1 in blue). An outline of the quadcopter is shown in green for reference.

Fig. 3 Bioinspired sensor module.

(A) Arrangement and placement of five paired pressure probes placed to maximize pressure deltas when close to surfaces. (B) Pressure sensor module components comprising the pressure sensor array, adapter printed circuit board, and microcontroller. (C) Schematic diagram showing internal routing tracks connecting paired probes (with Fore-Aft in green, Port-Starboard in yellow, ForwardPort-AftStarboard in dark blue, and ForwardStarboard-AftPort in orange, from top to bottom in light blue) to pressure sensors through a tube network shown in (D). (E) Free-flying prototype of a mosquito-inspired surface detection device. (F and G) Differential pressure delta with proximity to ground (F) and wall (G); shaded regions indicate 1 SD. Altitude is measured from the plane of the rotor hubs. Wall proximity is measured from the nearest rotor hub.

This simple model was able to detect both ground and wall effects. Pressure differential increases with surface proximity (Fig. 3, F and G) and with sufficient signal to provide alarm thresholds (tables S1 and S3) for each proximity condition. The complete module weighed just 9.2 g (see table S2 for a detailed mass breakdown).

The device successfully emulated the mosquito model behavior by identifying nearby obstacles during flight. Initially, the quadcopter was flown tethered (Fig. 4, A and B), was then piloted (Fig. 4C), and, finally, flew autonomously using positional feedback from a motion capture system. Ground (Fig. 4D and figs. S9 and S10) and wall (Fig. 4, E to G) planes could be discriminated using appropriately placed sensor combinations monitoring induced flow-field changes. Previous quadcopter studies have detected proximal surfaces by combining measured rotor speeds required for stable hovering with an aerodynamic model of the rotor and the motor speed required to support weight (15). Others have detected external flows such as fans emulating the downwash of another vehicle (16) or successfully incorporated flight dynamics models of the specific quadcopter platform and used them to infer obstacle proximity by the forces and torques acting on the vehicle (17). Our method requires no a priori aerodynamic or rigid body models to function, only basic thresholds. It is therefore a more direct measure of surface proximity and needs little or no processing to function.

Fig. 4 Demonstration of aerodynamic imaging in a quadcopter.

(A) Tethered wall proximity test with the wall on the forward side of the quadcopter. Yellow arrows indicate forward and aft red indicator lights. (B) Tethered ground proximity test. Yellow arrows indicate red alarm lights illuminating when ground is detected. (C) Piloted free-flight test of ground detection. (D to G) Long-exposure photographs of autonomous test of ground detection. (D) Oblique side view showing perpetual flight lights in blue and detection indicator lights in red. The ground was detected twice. (E to G) Top view of three wall detection trials. A single surface detection indicator light illuminates on one side nearest the wall before the quadcopter moves away from the obstruction. A strobe flash before the end of the exposure captures the quadcopter toward the end of its flight.

Supplementary Materials

Materials and Methods

Figs. S1 to S10

Tables S1 to S3

Caption for Movie S1

References (2022)

Movie S1

References and Notes

Acknowledgments: We thank the RVC flight group, P. Kimber and C. Jones of the Defense Science and Technology Laboratory (Dstl), and N. Short for discussions and H. Liu of Chiba University for permission to use the CFD simulator. Funding: This work was funded in part by the Autonomous Systems Underpinning Research (ASUR) program under the auspices of Dstl, the UK Ministry of Defense (R.J.B.), and the Biotechnology and Biological Sciences Research Council (BB/J001244/1 to R.J.B.). S.M.W. was supported by a Royal Society University Research Fellowship. Work at the University of Brighton was funded by the Medical Research Council (MR/N004299/1). P.S. was partly supported by a Rising Stars Award from the University of Brighton. Author contributions: R.J.B. conceived the experiments with T.N., N.P., S.M.W., and I.J.R. S.M.W., I.J.R., and P.S. advised on the experimental protocol. N.P. designed and built the quadcopter sensor module. N.P. and J.A.C. built the quadcopter module communication links. I.J.R. and P.S. gathered and processed data for the JO and femoral hairs. R.J.B., T.N., N.P., and S.M.W. wrote the manuscript. All authors contributed to editing the manuscript. Competing interests: Some of this work was used to support the filing of patent WO 2019/002892 A1. Data and materials availability: All data are available in the main text or the supplementary materials. Mosquito kinematics are available from Bomphrey et al. (13). The CFD solver (18) and kinematics acquisition code (19) are described in further detail elsewhere.

Stay Connected to Science

Navigate This Article