AV Capability Testing with Hyperplane Search Boundaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training and testing autonomous vehicles (AVs) in the physical world is costly, time-consuming, and poses challenges in simulating all driving scenarios, especially those that are dangerous or difficult to recreate.
Innovation Solution
The use of hyperplane searches in a simulation environment to efficiently define and test AV capabilities by representing vehicle capabilities with quantifiable metrics and sampling the N-dimensional parameter space sparsely to find pass/fail boundaries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If physical world testing is used to train and test autonomous vehicles, then realistic driving scenarios can be validated, but the process becomes costly and time-consuming
Solution Approach 1:
The patent creates virtual copies of physical driving environments, vehicles, and scenarios through high-fidelity simulation. The simulation system replicates real-world physics, sensor behaviors, and traffic patterns, allowing comprehensive testing without physical deployment. This copying approach maintains validation reliability while eliminating the time and cost constraints of physical testing.
Solution Approach 2:
The system performs preliminary testing and validation in the virtual environment before physical deployment. By pre-validating autonomous vehicle algorithms, edge cases, and safety protocols in simulation, the patent eliminates the need for extensive physical trial-and-error testing. This preliminary action in virtual space resolves the contradiction by providing reliable validation results without the time penalty of physical testing.
2Reliability
If comprehensive physical testing is conducted to cover all driving scenarios, then vehicle safety is improved, but the complexity and cost of testing increases significantly
Solution Approach 1:
The patent transitions the testing problem from physical space to virtual/digital space, adding a new dimension for conducting tests. This dimensional shift allows comprehensive scenario coverage including dangerous or rare edge cases that would be impractical to recreate physically. The virtual environment handles complexity through software parameterization rather than physical system complexity, resolving the contradiction between thorough safety validation and testing system complexity.
3Measurement precision
If dangerous or difficult driving scenarios are recreated in the physical world, then edge case validation is achieved, but the risk and cost of testing increase
Solution Approach 1:
The patent introduces a virtual simulation environment as an intermediary between the developer and dangerous testing scenarios. This intermediary allows precise recreation and testing of hazardous edge cases such as pedestrian crossings, intersection conflicts, and adverse weather conditions without exposing physical vehicles or personnel to actual danger. The simulation mediator maintains measurement precision for edge case detection while eliminating physical testing risks.
Data Source
AI summary
Systems and methods for measuring autonomous vehicle (AV) capabilities in a simulation environment are provided. A method includes selecting, by a computer-implemented system, a first plurality of sampling points in an N-dimensional parameter space associated with a vehicle capability test, classifying, by the computer-implemented system, each sampling point of the first plurality of sampling points into a first class based on a vehicle passing a test scenario associated with the respective sampling point; or a second class based on the vehicle failing the test scenario associated with the respective sampling point; and determining, by the computer-implemented system based on the classifying, a capability of the vehicle. The classifying can include computing a first hyperplane in the N-dimensional parameter space to separate the first plurality of sampling points into the first class and the second class.