Autonomous Driving Scenario Simulation for Edge-Case AI Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating training data for autonomous driving AI models are limited in securing data for edge cases, which are exceptional or extreme situations difficult to obtain, leading to performance limitations.
Innovation Solution
A method and apparatus that generate synthetic training data through autonomous driving simulation, using metadata and ground truth data from sensors to derive scenarios and create synthetic ground truth data for AI models, expanding the operational design domain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data is acquired through real-world autonomous driving, then the data reflects actual driving situations, but it is difficult to secure sufficient data for edge cases and requires enormous economic costs
Solution Approach 1:
The patent uses simulation technology to create virtual copies of real-world driving scenarios, particularly edge cases that are difficult to obtain in reality. By generating synthetic training data through simulated environments, the system can produce unlimited quantities of edge case data without requiring actual physical occurrences of these rare situations.
Solution Approach 2:
The system performs preliminary identification of edge cases through uncertainty assessment before actual data collection. By predicting which scenarios are likely to be edge cases in advance, the simulation can be configured to generate these specific scenarios beforehand, ensuring sufficient data availability for training.
2Measurement precision
If uncertainty-based selection is used to identify additional training data, then useful training examples can be found, but edge cases that are difficult to obtain remain insufficient
Solution Approach 1:
The patent introduces simulation technology as an intermediary between uncertainty assessment and data collection. The uncertainty assessment identifies potential edge cases, and the simulation acts as a mediator to generate the actual training data for these identified cases, bridging the gap between identification and data acquisition.
Solution Approach 2:
The system uses uncertainty assessment results as feedback to guide the simulation data generation process. By continuously monitoring model uncertainty and using this information to generate targeted synthetic data, the system creates a feedback loop that progressively improves coverage of edge cases.
3Productivity
If more real-world data is collected to cover all driving situations, then model performance can be improved, but economic costs increase enormously
Solution Approach 1:
Instead of collecting expensive real-world data for all possible driving situations, the patent creates virtual copies through simulation. This allows comprehensive coverage of diverse driving scenarios including rare edge cases without the proportional increase in data collection costs.
Solution Approach 2:
The system changes the parameter of data acquisition from physical collection to virtual generation. By transforming the data source from real-world sensors to simulation environments, the system maintains data diversity and model performance while dramatically reducing economic costs.
Data Source
AI summary
A method of generating training data for an artificial intelligence model may comprise: providing metadata and ground truth data generated from original data; generating accumulated data obtained by accumulating training results less than or equal to a standard indicator target value by comparing artificial intelligence model training results with a target indicator based on the ground truth data; deriving an autonomous driving scenario using the accumulated data; and generating synthetic ground truth data for training the artificial intelligence model by performing a simulation according to the autonomous driving scenario.