Simulation’s necessity in AV verification
Our approach to simulation within an AV verification framework
Our approach for the verification of AV decision making
Definition and creation of scenarios for simulation
Exploring the Future Potential of AI-Enabled Smartphone Processors
Abstract Simulation Scenario Generation for Autonomous Vehicle Verification
1. Abstract Simulation Scenario Generation
for Autonomous Vehicle Verification
Christopher Medrano-Berumen
and Dr. Mustafa Ilhan Akbas
Department of Computer Science
Advanced Mobility Institute
Florida Polytechnic University
2. About the Presentation
● Simulation’s necessity in AV verification
● Our approach to simulation within an AV verification
framework
● Our approach for the verification of AV decision making
● Definition and creation of scenarios for simulation
3. Introduction
● Proving autonomous vehicle safety will be essential for
public adoption
● Current methods of testing such as ‘shadow driving’ are
too costly and slow
● Autonomous vehicles deal with a complex operational
domain which needs to be perfectly simulated
4. Contributions
1. Defined a semantic language to create scenes
2. Developed a method to compose roadways
3. Defined a method to stitch road segments
4. Developed a self-verification method for road networks
7. Separation of Concerns
● Breaks down the autonomous vehicle into more easily
verifiable components
○ Perception
○ Object Recognition
○ Decision Making
● Verification of subsystems resembling that of hardware
verification
8. Decision Making
● Given an understanding of its surroundings, where the
autonomous vehicle decides its next action
● Verification will require defining what constitutes a good
decision
○ Vehicle follows local driving rules and traditional social rules
■ E.g. someone waving their hand at you to let you go ahead
● Need to appropriately prioritize rules versus safety
9. Decision Making (cont.)
● Newtonian Physics provide a
foundation for representing elements
● Assertions define the desired behavior
of the autonomous vehicle
● Design describes the properties of the
actual scenario
● Inputs & constraints describe what
was used to initialize the scenario
10. Simulation
● Simulation allows the brain of an autonomous vehicle to
be put in virtually any scenario
● Can run tests without safety risk of real-world testing
● Tests can be run in parallel around the clock
12. Choosing a Simulator
● Simulators span a wide-range of methods focusing on
different aspects of autonomous vehicles
○ I.e. machine learning, vehicle dynamics, etc.
● MATLAB provides a full-toolchain, Advanced Driving
Assistance System Toolbox
○ Integrates with the Simulink software, allowing modeling of different
components & vehicle in the loop
○ Does not simulate details outside of what is necessary for Newtonian
physics representation of objects
13. Matrix-based Semantic Language
● Breaks down road/scenario properties into generatable
features using matrices with numeric values
● Matrices passed into scenario generator that feeds the
appropriate information into the simulation environment
14. Semantic Language (cont.)
● Road matrix defines road layout
○ Each row represents a piece of the road and its properties
● Actor matrix defines other actors in scenario
○ Includes other vehicles, pedestrians, and other objects interacting with
AV
15. Road Piece Generation
● Road layout is broken down into common pieces
● Attributes of each type of piece is parametrized into
randomizable features
● Attributes get shared between pieces to limit input
● Currently, the multi-lane road and the 4-way intersection
are implemented
16. Geometric Primitives
● A road follows 3 types of
lines:
○ Line
○ Arc
○ Clothoid
■ Changes curvature at constant
rate, used for smooth turning
● Permutations in sets of 3 make
up all common forms of single
roads
17. Multi-lane Road Piece
● Constructed as a building block for
other roads
● Puts together a road from primitives
w/ all parameters in mind
○ Lanes
○ Length
○ Starting and ending Curvatures
○ Directions
○ Lane markers
18. Intersections
● Four-way Intersections
● Each road has individual properties
○ Lanes
○ Direction
● Turning lanes are determined by number
of available lanes in each direction
● Laid out according to all four widths
19. Stitching Roads Together
● Roads are stitched together by lining up tangent lines
○ Pieces are rotated and shifted to line up end/start points
○ Same is done for geometric pieces of multilane road
● Intermediary piece is created between pieces w/ different
lane numbers to account for changing road sizes
● Makes for continuous road that the ego vehicle can move
through
20. Road Validation
● A check is done on each
piece to check legality
● Not placing illegal road
pieces ensures the
scenario is not invalid
● Current check involves
not allowing separate
roads to intersect
21. Conclusion
● AV verification is critical and necessary
● We propose a semantic language to define scenarios
● Our simulation methodology composes all road
geometries based on road pieces/primitives, stitches
them together, and validates results
● Future work includes:
○ Definition of more road types and their properties
○ Integration with HiL simulation
○ Exporting scenarios to more common formats