Pseudo-Random UAV Test Generation Using Low-Fidelity Path Simulator

Anas Shrinah, Kerstin Eder · 2025

Read Pseudo-Random UAV Test Generation Using Low-Fidelity Path Simulator by Anas Shrinah, Kerstin Eder (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-24172.txt.

Introduction

Testing drone software in high-grade simulators is safe but slow, and Anas Shrinah and Kerstin Eder built a faster first pass in work from twenty twenty-five. Their maker builds random test flights with a rough, low-grade simulator. That rough tool runs about ten times faster than the full one.

The rough path tool is a trimmed form of a known autopilot simulator. Flights that look unsafe in the rough pass move on to the full simulator for a firm check. That two-step flow saves costly full runs for the cases that matter. The work suits drone teams that must test many flights on a tight budget of time and machines.

Catalog record

CREATORAnas Shrinah, Kerstin Eder
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE10 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-24172

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Pseudo-Random UAV Test Generation Using Low-Fidelity Path Simulator Thanks: This work was supported by the EPSRC under Grant EP/V050966/1. Anas Shrinah, Kerstin Eder Abstract Simulation-based testing provides a safe and cost-effective environment for verifying the safety of Uncrewed Aerial Vehicles (UAVs). However, simulation can be resource-consuming, especially when High-Fidelity Simulators (HFS) are used. To optimise simulation resources, we propose a pseudo-random test generator that uses a Low-Fidelity Simulator (LFS) to estimate UAV flight paths. Our pseudo-random test generator is the overall winner of the UAV testing competitions in SBFT and ICST 2025 1 1 1 https://github.com/skhatiri/UAV-Testing-Competition . This work simplifies the PX4 autopilot HFS to develop a LFS, which operates one order of magnitude faster than the HFS. Test cases predicted to cause safety violations in the LFS are subsequently validated using the HFS. I Introduction This paper presents PRGenUAV-LFS, a test generation tool that participated in the UAV Testing Competition organised by the Search-Based and Fuzz Testing (SBFT 2025) workshop [ 3 ] . This competition uses the Aerialist test bench to automate UAV software testing and simulation [ 2 ] . Participating test generators should create challenging scenarios by placing obstacles within a flying arena to force the UAV to fly dangerously close to them. A flight path is considered dangerous if the distance between the UAV and any obstacle is less than 1.5m at any time during the flight. The software under test is the PX4-Avoidance system, an open-source, vision-based autonomous obstacle avoidance system built on top of PX4-Autopilot. PX4 employs high-fidelity simulation platforms such as Gazebo to perform flight testing. High-Fidelity Simulators (HFS) are computationally expensive and require significant time to complete simulations. For instance, simulating a single UAV test case with Aerialist can take several minutes. To overcome this challenge, we propose to use a Low-Fidelity Simulator (LFS) to estimate flight paths. This LFS is faster than the PX4 HFS by one order of magnitude and estimates flight paths using the same planner used by the HFS. This planner is a local planner based on the 3D Vector Field Histogram (3DVFH*) algorithm [ 1 ] . Our test generator samples test cases from a predefined range to produce valid and interesting tests. These tests are evaluated using the LFS to predict the minimum distance between the UAV’s estimated path and the obstacles. Only tests with acceptable expected minimum distances are subsequently tested using the HFS. Fig. 1: Simulated test case example. II Pseudo-Random Test Generator Test cases are generated for flying missions; each mission consists of a start point, a series of waypoints, and a landing point. This test generation tool identifies the flight segment closest to the middle line of the Obstacle Designated Arena (ODA) that intersects with the top and bottom borders of the ODA. After that, the tool places the obstacle to obstruct this segment of interest (SoI). Obstacle configurations are determined based on the slope and the direction of the SoI. This section details the process of generating test scenarios for the case shown in Figure 1 . Other cases (different slops and directions of the SOI) are handled by applying geometric transformations to the method described below. Obstacles are cuboids defined by the coordinates of the centre, length, and width of its base. We fixed the obstacle height to 20 m, as the UAV is restricted from flying above obstacles. For simplicity, we refer to an obstacle by its base We observed that only two obstacles are required to compel the UAV to follow an S-shaped path, which can cause the UAV to fly close to the obstacles. We chose to set the the width of the bases of the two obstacles to 2m. The test generator randomly selects the diagonal length, rotation angle, and y-coordinate of the base of the first obstacle from predefined ranges to ensure the the base is fully contained within the designated area. The base length is then calculated using the diagonal and the width, and the x-coordinate is determined using the y-coordinate and the equation of the line defined by the SoI. The first obstacle must be placed in the path of the UAV to direct it away from the goal. This obstacle should intersect the SoI at an obtuse angle and at a point that divides the length of its base into two parts, A and B, as illustrated in Figure 1 . The length of part A should be one-third of the base's length, while part B spans the remaining two-thirds. This arrangement causes the UAV to track the obstacle along part A, as this path aligns the UAV’s heading closer to the goal direction. This behaviour is enforced by the planner because its cost function favours path nodes with orientations close to the goal direction. In Figure 1 , part A (B) is located to the left (right) of the SoI. To ensure proper placement, one-third (two-thirds) of the diagonal length must be less than the distance between the SoI and the ODA's left (right) border. This constraint guarantees that, even if the diagonal is aligned parallel to the X-axis, the vertices of the diagonal remain within the ODA, ensuring the obstacle is horizontally contained. To ensure vertical containment of the obstacle within the ODA, the y-coordinate must be chosen such that, even if the diagonal is aligned parallel to the Y-axis, the vertices of the diagonal remain inside the ODA. This is achieved by constraining the y-coordinate plus (minus) half of the diagonal must be less (greater) than the y-coordinate of the ODA's top (bottom) border. These constraints consider only the two extreme cases, when the diagonal is parallel to the X or Y axis, because these cases encompass all possible obstacle rotations, ensuring that the entire obstacle remains within the ODA under any orientation. After flying near the first obstacle, the planner will adjust the UAV's path to realign its orientation towards the goal (the end of the SoI). To prevent the UAV from flying away from the first obstacle when correcting its path, the second obstacle is placed perpendicular to the first obstacle and close enough to the first obstacle, but not too close to risk the planner failing to find a feasible solution. The x- and y-coordinates of the centre of the second obstacle's base are specified relative to the first obstacle to achieve this requirement and to the borders of the ODA to avoid having parts of this obstacle outside the ODA. The length of the second obstacle needs to be long enough to prevent the UAV from circling it from below; it is set to 1.75 of the first obstacle length. The next section describes the LFS used for evaluating the test cases generated. III low-fidelity path simulator The 3DVFH* algorithm uses 3D point clouds from a stereo camera to plan obstacle-avoiding paths. While an HFS generates these point clouds in PX4 simulation, our LFS leverages the Open3D library to produce them [ 4 ] . The process begins by creating a scene that contains a floor plane and the two given obstacles. The scene is then rendered into a depth image relative to the UAV position and orientation. Finally, a 3D point cloud is

Read the complete text