Verification of Autonomous Vehicles Through Simulation
Using MATLAB ADAS Toolbox
Christopher Medrano-Berumen and Dr. Mustafa İlhan Akbaş
Department of Computer Science, Florida Polytechnic University
Research Objectives
• Create a standardized verification framework
that can be used by industry and regulators to
evaluate autonomous vehicles for safety
• Develop a model of computation that will
integrate into the simulation to provide
coverage and repeatability of test scenarios
• Create a semantic language to define scenes
when generating scenarios
Motivation
Autonomous Vehicle (AV) technology is expected to
have a disruptive effect on industry. However, there
is still no reliable and standard method to verify an
AV’s decision taking process. The shadow driving,
which has been a majority of the real life testing so
far, will take one trillion miles and cost over $300
billion if it is to happen at all [1]. Hence, verifying
autonomous vehicles in an efficient and reliable way
is essential if the public are to be willing to accept
this new technology.
Introduction
We are part of the Advanced Mobility Institute at
Florida Polytechnic University. For the verifica-
tion of autonomous vehicles, we are inspired by the
successful testing methodologies used in hardware
verification and our main focus is applying these
methodologies for the verification of AV decision tak-
ing process through modeling and simulation.
Resources
The following software and add-ons were required to
complete the research:
• MATLAB R2018b
• MATLAB Advanced Driving Assistance System
Toolbox
Framework Architecture
The simulation framework begins in the main pro-
gram which will have certain tuning options before
starting the simulations. The scenarios are then ran-
domly generated according to those inputs and fed
into the actual simulation. Here, the scenario is gen-
erated, the ego vehicle is placed into it, and each step
is run with the ego vehicle responding to it through
the entire cycle of sensing, making a decision, and
acting on that decision.
Figure 1: Simulation Framework Architecture
Data is then collected from the scenario, that is,
whether each decision was safe, good, or neither,
and the scenario’s matrix. Because the scenario was
generated by that matrix, knowing the matrix means
that the scenario can then be recreated whenever.
Figure 2: Scenario where ego vehicle is leading a platoon
Scenario Generation
Semantic Language: The first phase in our ap-
proach is the creation of a semantic language to de-
scribe driving scenarios. This language, read in as a
matrix where each row is a different assertion, can
then be parsed to generate the scenario. To do this,
road network and actors are reduced to their most
basic elements which are then parametrized.
Road Topology Generation: Putting the roads
in the simulation requires the points it follows and
the width, so road generation is a matter of taking
input and generating the points that follow the road
described. This is done differently according to the
geometric primitives that form the road pieces. The
goal in defining the different road pieces is to con-
strain the generated scenario to realistic road net-
works and situations.
Figure 3: Roads created by generated points and driving paths
Actor Generation: Besides the ego vehicle, ac-
tors in the scenario are generated randomly and are
also comprised of different types. This includes vehi-
cles which can be cars, trucks, etc., pedestrians, and
so on. The different behaviors that they may take
in the scenario are also programmed in. These in-
clude crossing the road for pedestrians and swerving
outside of the lane for vehicles.
Test Scenario Generation: By pseudo-random
input generation for all of the created pieces, one
can begin to test whether the autonomous vehicle
will take the correct decision in any situation.
Research Results
• Developed a semantic language for the generation
of AV verification scenarios
• Implemented the simulation framework for
scenario generation in MATLAB
• Extended the road and actor generation
properties of MATLAB ADAS toolbox
• Developed a validity check for generated scenarios
Future Research Agenda
• Improving the semantic language and extending
road piece and traffic scenario database
• Integrating a model of computation into the AV
decision step for coverage analysis
• Enabling scenario-export to formats that would
permit cross-platform simulation, e.g.
OpenDRIVE
References
[1] Nidhi Kalra and Susan M. Paddock.
Driving to safety: How many miles of driving would it take
to demonstrate autonomous vehicle reliability?
2016.
[2] Inc. The MathWorks.
MATLAB and Automated Driving System Toolbox,
Release R2018b.
Natick, Massachusetts, United States.
Contact Information
• Web: https://floridapoly.edu/advanced-
mobility-institute
• Email: cmedranoberumen2844@floridapoly.edu
• Email: makbas@floridapoly.edu
• Phone: +1 (863) 874 8546

Verification of Autonomous Vehicles Through Simulation Using MATLAB ADAS Toolbox

  • 1.
    Verification of AutonomousVehicles Through Simulation Using MATLAB ADAS Toolbox Christopher Medrano-Berumen and Dr. Mustafa İlhan Akbaş Department of Computer Science, Florida Polytechnic University Research Objectives • Create a standardized verification framework that can be used by industry and regulators to evaluate autonomous vehicles for safety • Develop a model of computation that will integrate into the simulation to provide coverage and repeatability of test scenarios • Create a semantic language to define scenes when generating scenarios Motivation Autonomous Vehicle (AV) technology is expected to have a disruptive effect on industry. However, there is still no reliable and standard method to verify an AV’s decision taking process. The shadow driving, which has been a majority of the real life testing so far, will take one trillion miles and cost over $300 billion if it is to happen at all [1]. Hence, verifying autonomous vehicles in an efficient and reliable way is essential if the public are to be willing to accept this new technology. Introduction We are part of the Advanced Mobility Institute at Florida Polytechnic University. For the verifica- tion of autonomous vehicles, we are inspired by the successful testing methodologies used in hardware verification and our main focus is applying these methodologies for the verification of AV decision tak- ing process through modeling and simulation. Resources The following software and add-ons were required to complete the research: • MATLAB R2018b • MATLAB Advanced Driving Assistance System Toolbox Framework Architecture The simulation framework begins in the main pro- gram which will have certain tuning options before starting the simulations. The scenarios are then ran- domly generated according to those inputs and fed into the actual simulation. Here, the scenario is gen- erated, the ego vehicle is placed into it, and each step is run with the ego vehicle responding to it through the entire cycle of sensing, making a decision, and acting on that decision. Figure 1: Simulation Framework Architecture Data is then collected from the scenario, that is, whether each decision was safe, good, or neither, and the scenario’s matrix. Because the scenario was generated by that matrix, knowing the matrix means that the scenario can then be recreated whenever. Figure 2: Scenario where ego vehicle is leading a platoon Scenario Generation Semantic Language: The first phase in our ap- proach is the creation of a semantic language to de- scribe driving scenarios. This language, read in as a matrix where each row is a different assertion, can then be parsed to generate the scenario. To do this, road network and actors are reduced to their most basic elements which are then parametrized. Road Topology Generation: Putting the roads in the simulation requires the points it follows and the width, so road generation is a matter of taking input and generating the points that follow the road described. This is done differently according to the geometric primitives that form the road pieces. The goal in defining the different road pieces is to con- strain the generated scenario to realistic road net- works and situations. Figure 3: Roads created by generated points and driving paths Actor Generation: Besides the ego vehicle, ac- tors in the scenario are generated randomly and are also comprised of different types. This includes vehi- cles which can be cars, trucks, etc., pedestrians, and so on. The different behaviors that they may take in the scenario are also programmed in. These in- clude crossing the road for pedestrians and swerving outside of the lane for vehicles. Test Scenario Generation: By pseudo-random input generation for all of the created pieces, one can begin to test whether the autonomous vehicle will take the correct decision in any situation. Research Results • Developed a semantic language for the generation of AV verification scenarios • Implemented the simulation framework for scenario generation in MATLAB • Extended the road and actor generation properties of MATLAB ADAS toolbox • Developed a validity check for generated scenarios Future Research Agenda • Improving the semantic language and extending road piece and traffic scenario database • Integrating a model of computation into the AV decision step for coverage analysis • Enabling scenario-export to formats that would permit cross-platform simulation, e.g. OpenDRIVE References [1] Nidhi Kalra and Susan M. Paddock. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? 2016. [2] Inc. The MathWorks. MATLAB and Automated Driving System Toolbox, Release R2018b. Natick, Massachusetts, United States. Contact Information • Web: https://floridapoly.edu/advanced- mobility-institute • Email: cmedranoberumen2844@floridapoly.edu • Email: makbas@floridapoly.edu • Phone: +1 (863) 874 8546