Autonomous Vehicle Rare Scenario Simulation for ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in capturing sufficient data for training machine learning models to handle rare scenarios, which occur infrequently during daily driving.
Innovation Solution
An AV management system determines a target likelihood value range for simulating rare scenarios, generates new scenarios by modifying existing ones, and uses machine learning models to assess the likelihood and relevance of these scenarios, thereby prioritizing relevant rare scenarios for simulation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If autonomous vehicles capture more real-world data to train machine learning models, then the model performance improves, but the time and resources required increase significantly
Solution Approach 1:
The system performs preliminary analysis of available AV scene data to identify rare scenarios before simulation. By determining likelihood values and prioritizing scenarios in advance, the system prepares training data efficiently without requiring extensive real-world data collection, thus improving model performance while reducing time loss.
Solution Approach 2:
Instead of collecting more real-world data, the system creates synthetic copies of rare scenarios through computer-generated simulations. These simulated AV scene data serve as artificial training samples that replicate rare events without requiring actual real-world occurrences, thereby improving model training efficiency while reducing time and resource requirements.
2Adaptability or versatility
If autonomous vehicles simulate more rare scenarios, then the coverage of edge cases improves, but the computing resources required increase
Solution Approach 1:
The system changes the parameter of scenario selection by using likelihood values to prioritize which rare scenarios to simulate. Instead of uniformly simulating all possible rare scenarios, the system adjusts parameters to focus computational resources on scenarios with likelihood values indicating they are most relevant and likely to occur, thereby improving scenario coverage while reducing computing resource consumption.
Solution Approach 2:
The system uses machine learning models to generate likelihood values for different scenarios based on available AV scene data. This feedback mechanism allows the system to identify which rare scenarios are most worthy of simulation, creating a closed-loop process that optimizes computing resource allocation by continuously evaluating and prioritizing scenarios based on their relevance and probability.
3Reliability
If the system simulates all possible rare scenarios, then comprehensive training is achieved, but the processing time increases
Solution Approach 1:
The system applies partial action by selectively simulating only the most relevant rare scenarios rather than all possible scenarios. By using likelihood values to determine which scenarios warrant simulation, the system performs sufficient training on high-priority scenarios while omitting low-priority ones, achieving acceptable training comprehensiveness with significantly improved processing efficiency.
Solution Approach 2:
The system performs preliminary filtering and prioritization of rare scenarios using machine learning models before simulation. By pre-processing the scenario selection based on likelihood values derived from available AV scene data, the system identifies a focused subset of scenarios that require simulation, thereby maintaining training comprehensiveness while dramatically reducing processing time and resource requirements.
Data Source
AI summary
Systems and techniques are provided for rare scenario handling in autonomous vehicles. In some aspects, an AV management system can determine a target likelihood value range for identifying relevant rare scenarios. In some cases, the AV management system may use the target likelihood value range to generate new relevant rare scenarios that have a likelihood value that falls within the target likelihood value range. In some examples, the AV management system may initiate simulations of the generated relevant rare scenarios and capture synthetic AV scene data, which can be used to train machine learning models used in AVs. As a result, the performance of the AVs may be improved when faced with rare scenarios.