Skip to content
This repository has been archived by the owner on Sep 22, 2023. It is now read-only.
/ tool-competition-av Public archive

The repository hosts the code for the SBST CPS Tool competition for testing autonomous cars.

License

Notifications You must be signed in to change notification settings

se2p/tool-competition-av

Repository files navigation

Cyber-Physical Systems Testing Tool Competition [OUTDATED]

This repository refers to the 2022 edition, to check the latest version and join the current competition, you can visit the current repository

Contacts

For more information on the 2022 edition, contact:

Dr. Alessio Gambi - IMC Krems, Austria

Dr. Vincenzo Riccio - Università di Udine, Italy

Goal

The SBST Workshop offers a challenge for software testers who want to work with self-driving cars in the context of the usual tool competition.

The competitors should generate virtual roads to test a lane keeping assist system using the provided code_pipeline.

The generated roads are evaluated in the BeamNG.tech driving simulator. This simulator is ideal for researchers due to its state-of-the-art soft-body physics simulation, ease of access to sensory data, and a Python API to control the simulation.

Video by BeamNg GmbH

Note: BeamNG GmbH, the company developing the simulator, kindly offers it for free for researcher purposes upon registration (see Installation).

Comparing the Test Generators

Deciding which test generator is the best is far from trivial and, currently, remains an open challenge. In this competition, we rank test generators by considering various metrics of effectiveness and efficiency that characterize the generated tests but also the process of generating them, i.e., test generation. We believe that our approach to compare test generators is objective and fair, and it can provide a compact metric to rank them.

Ranking Formula

The formula to rank test generators is the following weighted sum:

rank = a * OOB_Coverage + b * test_generation_efficiency + c *  test_generation_effectiveness

where:

  • OOB_Coverage captures the effectiveness of the generated tests that must expose as many failures as possible (i.e., Out Of Bound episodes) but also as many different failures as possible. We compute this metric by extending the approach adopted in the previous edition of the competition with our recent work on Illumination Search. As an example, our novel approach has been already adopted for the generation of relevant test cases from existing maps (see SALVO). Therefore, we identify tests' portion relevant to the OOBs, extract their structural and behavioral features, and populate feature maps of a predefined size (i.e., 25x25 cells). Finally, we define OOB_Coverage by counting the cells in the map covered by the exposed OOBs. Larger values of OOB_Coverage identify better test generators.

  • test_generation_efficiency captures the efficiency in generating, but not executing, the tests. We measure it as the inverse of the average time it takes for the generators to create the tests normalized using the following (standard) formula:

    norm(x) = (x - min) / (max - min)

    Where min and max are values empirically found during the benchmarking as the minimum and maximum average times for generating test across all the competitors.

  • test_generation_effectiveness captures the ability of the test generator to create valid tests; therefore, we compute it as the ratio of valid tests over all the generated tests.

Setting the Weights

We set the values of the in the ranking formula's weights (i.e., a, b, and c) to rank higher the test generators that trigger many and different failures; test generation efficiency and effectiveness are given equal but secondary importance. The motivation behind this choice is that test generators' main goal is to trigger failures, while being efficient and effective in generating the tests is of second order importance.

The following table summarizes the proposed weight assignment:

a b c
0.6 0.2 0.2

Implement Your Test Generator

We make available a code pipeline that will integrate your test generator with the simulator by validating, executing and evaluating your test cases. Moreover, we offer some sample test generators to show how to use our code pipeline.

Information About the Competition

More information can be found on the SBST tool competition website: https://sbst22.github.io/tools/

Repository Structure

Code pipeline: code that integrates your test generator with the simulator

Self driving car testing library: library that helps the integration of the test input generators, our code pipeline, and the BeamNG simulator

Scenario template: basic scenario used in this competition

Documentation: contains the installation guide, detailed rules of the competition, and the frequently asked questions

Sample test generators: sample test generators already integrated with the code pipeline for illustrative purposes

Requirements: contains the list of the required packages.

License

The software we developed is distributed under GNU GPL license. See the LICENSE.md file.