Research ArticleSENSORS

Robotic ecology: Tracking small dynamic animals with an autonomous aerial vehicle

See allHide authors and affiliations

Science Robotics  17 Oct 2018:
Vol. 3, Issue 23, eaat8409
DOI: 10.1126/scirobotics.aat8409
  • Fig. 1 The aerial robot system is designed to track small animals with lightweight radio tags attached.

    Swift parrots (L. discolor) (A) are considered here; noisy miners (M. melanocephala) (B) were considered in our previous work (22). (C) The robot was able to track swift parrots and yielded performance comparable with an expert human operator performing traditional wildlife telemetry tracking. The multirotor platform, shown from the front (D) and rear (E), includes a lightweight directional antenna system and payload that receives the signal strength from the tag (22). These data were then transmitted to a GCS for processing and online decision-making.

  • Fig. 2 Obtaining range-azimuth likelihood functions from observations.

    (Top row) Two example observations taken online with a stationary target. The radial plots illustrate real RSSI readings (green line) g and a third-order Fourier series model φ(θ) of the radiation pattern (black line). The model is offset (rotated) such that it is oriented toward the true bearing to the target θ*, and the RSSI values are offset by the maximum correlation μΘ(g) = arg maxθ rϕ(θ),g. These offsets are illustrated with dashed green and black radial lines. (Left) The maximum value correlation coefficient Embedded Image maps to a bearing-error Embedded Image, which is illustrated in the grid plots below. (Right) The maximum RSSI value gmax maps to an expected range μR(g) with a fixed range-error Embedded Image, giving the associated grid plots below.

  • Fig. 3 Example of our Bayesian data fusion method to obtain target estimates.

    The distributions shown are spatially discrete grids over a 750-m2 area (with grid lines every 100 m for illustrative purposes only). (Left to right) The bearing-only likelihood function ℓΘ, the range-only likelihood function R, the combined likelihood function ℓ, and the posterior belief p(·). In each column, the first observation (k = 1) is shown in the bottom grid, the last observation (k = 4) is at the top, and higher probability mass is represented as darker, raised regions. The UAV location xk is indicated by a green dot, the target location Embedded Image in purple, and the maximum likelihood estimate ŷk in yellow.

  • Fig. 4 Evaluating the performance of the robotic system through comparison with human trackers.

    We performed two flights at each of four trial sites (eight flights in total). The box plot illustrates the estimate errors (on the y axis) for both the robot (green and blue) and the human (white and gray) trackers as a function of the observation number (on the x axis). The blue boxes labeled “Robot (certain)” indicate scenarios where the bird remained stationary during the trial and the final location was known.

  • Fig. 5 Recorded spatial distribution of swift parrots.

    A map of southeast Australia, where a majority of swift parrots have been sighted. Inset: Heatmap illustrating aggregated posterior distributions from our trial in July 2017. The posterior distributions of all trials were normalized and aggregated to give an indication of the most likely foraging and roosting areas. White symbols indicate locations where ground truth data (confirmed by visual sightings) were available; each tag has a different symbol: ○, ×, +, or *. Map data: Google, DigitalGlobe, CNES/Airbus.

Supplementary Materials

  • Supplementary Materials

    The PDF file includes:

    • Table S1. Field trial data.

    Download PDF

    Other Supplementary Material for this manuscript includes the following:

    • Movie S1 (.mp4 format). Flight demonstration.

    Files in this Data Supplement:

Stay Connected to Science Robotics

Navigate This Article