Autonomous Driving Scenario Embeddings for Duplicate Simulation Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle (AV) training methods rely heavily on real-world testing, which may not expose AVs to every possible scenario, leading to challenges in scaling AV fleets and handling edge cases, and simulating large numbers of duplicate scenarios consumes significant computing resources without meaningful improvement.
Innovation Solution
A simulation environment using a scenario generator and machine learning-based embedding model to filter and select unique driving scenarios, employing a graphical processing unit (GPU) for parallel processing to compare scenarios and reduce duplicates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real-world road tests are used to train AV software, then the AV can learn from actual driving conditions, but it cannot encounter every possible scenario including rare edge cases
Solution Approach 1:
The patent creates virtual copies of real-world driving scenarios through simulation environments. By capturing actual driving data and reconstructing scenarios in a virtual setting, the system can replicate rare edge cases and uncommon situations that would be difficult to encounter in real-world testing, thereby improving both reliability through realistic training and adaptability through expanded scenario coverage.
Solution Approach 2:
The system performs preliminary analysis of driving scenarios to identify rare and unusual cases before conducting full simulation training. By pre-screening scenarios using embedding models and similarity comparisons, the system prepares and selects only the most valuable training cases in advance, making the subsequent simulation process more efficient and targeted toward improving software reliability for edge cases.
2Adaptability or versatility
If simulations are used to train AV for all possible scenarios, then scenario coverage is improved, but computing resources are wasted on simulating similar duplicate scenarios
Solution Approach 1:
The patent introduces embedding models as an intermediary layer between scenario generation and simulation execution. These embeddings compress complex scenario data into compact representations that can be quickly compared for similarity. This intermediary filtering mechanism identifies and eliminates duplicate scenarios before they consume expensive simulation computing resources, thereby maintaining comprehensive scenario coverage while significantly reducing energy loss.
Solution Approach 2:
Instead of simulating all generated scenarios, the system performs a partial action by selectively simulating only those scenarios that are sufficiently different from previously processed ones. The embedding-based filtering performs an excessive preliminary comparison to ensure no valuable unique scenarios are missed, then restricts full simulation to only the necessary subset, optimizing the balance between coverage and resource efficiency.
3Reliability
If all generated scenarios are simulated, then training comprehensiveness is improved, but processing time and computational cost increase significantly
Solution Approach 1:
The patent segments the scenario processing pipeline into distinct stages: scenario generation, embedding creation, similarity comparison, and selective simulation. By dividing the workflow and inserting the embedding-based filtering segment, the system maintains training comprehensiveness through multi-stage processing while reducing overall processing time by eliminating redundant scenarios before the computationally intensive simulation phase.
Data Source
AI summary
A computer system for filtering autonomous driving scenarios for simulations includes circuitry and a graphical processing unit (GPU). The circuitry receives data encoding a first autonomous driving scenario, and generates a first embedding of the first autonomous driving scenario based on a model for transforming autonomous driving scenario data to an embedding. A GPU compares the first embedding to a test bank of embeddings representing other autonomous driving scenarios. If the first embedding has at least a threshold similarity to other embeddings in the test bank, the first autonomous driving scenario is discarded from a simulation pipeline for simulating autonomous driving scenarios.