Eagle Strategy with Local Search
for Scenario Based Validation of
Autonomous Vehicles
Quentin Goss and Mustafa Ilhan Akbas
Department of Electrical Engineering and Computer Science
Embry-Riddle Aeronautical University
Daytona Beach, Florida
gossq@my.erau.edu, akbasm@erau.edu
Outline
• Introduction
• Eagle Strategy with Local Search
• Estimating Edge Scenario Boundaries
• Demonstration
• Conclusion
Introduction
Introduction
Previous Research
• Foretify + MSDL by Foretellix
• Beta Testing
• Modeling AV scenarios based on real-life CAV Accidents
• accident report --> scenario --> simulation
• General-case + modular
• IEEE IV 2021
Introduction
Previous Research
• Foretify + MSDL by Foretellix
• Beta Testing
• Modeling AV scenarios based on real-life CAV Accidents
• accident report --> scenario --> simulation
• General-case + modular
• IEEE IV 2021
• Observations
• Black box scenarios
• Parameters + Outcome
• Run in batches
• Lots of them are uneventful/uninteresting.
Introduction
Previous Research
• Foretify + MSDL by Foretellix
• Beta Testing
• Modeling AV scenarios based on real-life CAV Accidents
• accident report --> scenario --> simulation
• General-case + modular
• IEEE IV 2021
• Observations
• Black box scenarios
• Parameters + Outcome
• Run in batches
• Lots of them are uneventful/uninteresting.
How can we build a testing regime
which selects scenario parameters
in such a way that prioritizes
finding interesting scenarios first?
Introduction
• Scope/Focus of Research
• Autonomous Vehicles (AV) Scenarios
• Complex AV Scenarios
• Blackbox AV Scenarios
Introduction
• Scope/Focus of Research
• Autonomous Vehicles (AV) Scenarios
• Complex AV Scenarios
• Blackbox AV Scenarios
• AV Scenarios have many parameters.
• Complexity increases with realism.
Introduction
• Scope/Focus of Research
• Autonomous Vehicles (AV) Scenarios
• Complex AV Scenarios
• Blackbox AV Scenarios
• AV Scenarios have many parameters.
• Complexity increases with realism.
• How to choose scenarios?
• Interesting
• Efficient
• How to process/visualize it?
Eagle Strategy with Local Search
Choosing Scenarios
with Eagle Strategy with Local Search
• Eagle Strategy (Yang and Deb, 2010)
• Stochastic Optimization
• Levy Walk + Firefly Algorithm
• Eagles' search for food.
Choosing Scenarios
with Eagle Strategy with Local Search
Choosing Scenarios
with Eagle Strategy with Local Search
Choosing Scenarios
with Eagle Strategy with Local Search
Choosing Scenarios
with Eagle Strategy with Local Search
Choose a point/parameters
far away from the last.
Choosing Scenarios
with Eagle Strategy with Local Search
Choose a point/parameters
far away from the last.
Choose a point/parameters
near the last known bug.
Parameter Selection
and representing scenarios by their parameters.
• Representation by parameters.
• How are scenarios parameters selected currently?
• Exhaustively
• Randomly
• Quasi-randomly
Parameter Selection
and representing scenarios by their parameters.
• Representation by parameters.
• How are scenarios parameters selected currently?
• Exhaustively
• Randomly
• Quasi-randomly
But we want to find the interesting scenarios!
• Bug Scenarios
Estimating Edge Scenario Boundaries
• Data from executed scenarios:
• Input parameters
• Outcome
• Properties
• N-dimensional
• Locations of bug scenarios unknown.
Estimating Edge Scenario Boundaries
• Data from executed scenarios:
• Input parameters
• Outcome
• Properties
• N-dimensional
• Locations of bug scenarios unknown.
Estimating Edge Scenario Boundaries
Review of data collected from.
1. How to visualize?
2. What data structure?
3. How can we estimate where
the edge scenario boundaries
are?
Estimating Edge Scenario Boundaries
The Voronoi Diagram
Estimating Edge Scenario Boundaries
The Graph
Demonstration
Demonstration
Scenario Design + Bug Classification
• Car 1 begins [0..1000]m from node C on edge 1 and
has a constant speed of [1..100]kph.
• Car 2 begins [0..1000]m from node C on edge 2 and
has a constant speed of [1..100]kph.
Demonstration
Comparison of Approaches
Demonstration
Visualizing the Bug Profile
Conclusion
Conclusion
• Blackbox Approach for AV Scenario Generation
• Parameter Range Sampling
• Efficiently prioritizes "interesting scenarios"
• Better than Random or Quasi-random range sampling
• More than double!
• Bug Profile Generation
• Voronoi Diagram
• Graph
• Scalable to N-dimensions!
Questions ?

Goss_ICCVE 2022.pdf

  • 1.
    Eagle Strategy withLocal Search for Scenario Based Validation of Autonomous Vehicles Quentin Goss and Mustafa Ilhan Akbas Department of Electrical Engineering and Computer Science Embry-Riddle Aeronautical University Daytona Beach, Florida gossq@my.erau.edu, akbasm@erau.edu
  • 2.
    Outline • Introduction • EagleStrategy with Local Search • Estimating Edge Scenario Boundaries • Demonstration • Conclusion
  • 3.
  • 4.
    Introduction Previous Research • Foretify+ MSDL by Foretellix • Beta Testing • Modeling AV scenarios based on real-life CAV Accidents • accident report --> scenario --> simulation • General-case + modular • IEEE IV 2021
  • 5.
    Introduction Previous Research • Foretify+ MSDL by Foretellix • Beta Testing • Modeling AV scenarios based on real-life CAV Accidents • accident report --> scenario --> simulation • General-case + modular • IEEE IV 2021 • Observations • Black box scenarios • Parameters + Outcome • Run in batches • Lots of them are uneventful/uninteresting.
  • 6.
    Introduction Previous Research • Foretify+ MSDL by Foretellix • Beta Testing • Modeling AV scenarios based on real-life CAV Accidents • accident report --> scenario --> simulation • General-case + modular • IEEE IV 2021 • Observations • Black box scenarios • Parameters + Outcome • Run in batches • Lots of them are uneventful/uninteresting. How can we build a testing regime which selects scenario parameters in such a way that prioritizes finding interesting scenarios first?
  • 7.
    Introduction • Scope/Focus ofResearch • Autonomous Vehicles (AV) Scenarios • Complex AV Scenarios • Blackbox AV Scenarios
  • 8.
    Introduction • Scope/Focus ofResearch • Autonomous Vehicles (AV) Scenarios • Complex AV Scenarios • Blackbox AV Scenarios • AV Scenarios have many parameters. • Complexity increases with realism.
  • 9.
    Introduction • Scope/Focus ofResearch • Autonomous Vehicles (AV) Scenarios • Complex AV Scenarios • Blackbox AV Scenarios • AV Scenarios have many parameters. • Complexity increases with realism. • How to choose scenarios? • Interesting • Efficient • How to process/visualize it?
  • 10.
    Eagle Strategy withLocal Search
  • 11.
    Choosing Scenarios with EagleStrategy with Local Search • Eagle Strategy (Yang and Deb, 2010) • Stochastic Optimization • Levy Walk + Firefly Algorithm • Eagles' search for food.
  • 12.
    Choosing Scenarios with EagleStrategy with Local Search
  • 13.
    Choosing Scenarios with EagleStrategy with Local Search
  • 14.
    Choosing Scenarios with EagleStrategy with Local Search
  • 15.
    Choosing Scenarios with EagleStrategy with Local Search Choose a point/parameters far away from the last.
  • 16.
    Choosing Scenarios with EagleStrategy with Local Search Choose a point/parameters far away from the last. Choose a point/parameters near the last known bug.
  • 17.
    Parameter Selection and representingscenarios by their parameters. • Representation by parameters. • How are scenarios parameters selected currently? • Exhaustively • Randomly • Quasi-randomly
  • 18.
    Parameter Selection and representingscenarios by their parameters. • Representation by parameters. • How are scenarios parameters selected currently? • Exhaustively • Randomly • Quasi-randomly But we want to find the interesting scenarios! • Bug Scenarios
  • 20.
  • 21.
    • Data fromexecuted scenarios: • Input parameters • Outcome • Properties • N-dimensional • Locations of bug scenarios unknown. Estimating Edge Scenario Boundaries
  • 22.
    • Data fromexecuted scenarios: • Input parameters • Outcome • Properties • N-dimensional • Locations of bug scenarios unknown. Estimating Edge Scenario Boundaries Review of data collected from. 1. How to visualize? 2. What data structure? 3. How can we estimate where the edge scenario boundaries are?
  • 23.
    Estimating Edge ScenarioBoundaries The Voronoi Diagram
  • 24.
    Estimating Edge ScenarioBoundaries The Graph
  • 25.
  • 26.
    Demonstration Scenario Design +Bug Classification • Car 1 begins [0..1000]m from node C on edge 1 and has a constant speed of [1..100]kph. • Car 2 begins [0..1000]m from node C on edge 2 and has a constant speed of [1..100]kph.
  • 27.
  • 28.
  • 29.
  • 30.
    Conclusion • Blackbox Approachfor AV Scenario Generation • Parameter Range Sampling • Efficiently prioritizes "interesting scenarios" • Better than Random or Quasi-random range sampling • More than double! • Bug Profile Generation • Voronoi Diagram • Graph • Scalable to N-dimensions!
  • 31.