Closed-Course Scenario Replay for Rare AV Training Cases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in handling infrequently encountered scenarios due to insufficient training data, which can lead to unsafe or abnormal responses.
Innovation Solution
A closed course environment is used to recreate important scenarios, allowing autonomous vehicles to interact with staged scenarios to gather additional training data, which is then used to adjust machine learning models, enhancing their ability to handle these rare situations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If autonomous vehicles are tested in real-world environments to improve machine learning models, then the models can learn from diverse scenarios, but infrequently encountered scenarios are insufficiently represented due to their rarity
Solution Approach 1:
The system performs preliminary identification of important but infrequently encountered scenarios from real-world testing data, then proactively recreates these scenarios in closed course environments before they are sufficiently represented in training data. This advance preparation ensures rare scenarios are adequately covered in model training.
Solution Approach 2:
The system creates simplified copies or representations of complex real-world scenarios by staging them in controlled closed course environments. These staged scenarios capture the essential characteristics of rare events while allowing repeated testing and data collection without the constraints of real-world conditions.
2Measurement precision
If more training data is collected from real-world scenarios, then model accuracy improves, but the time and resources required for data collection increase
Solution Approach 1:
Instead of collecting additional real-world data over extended periods, the system creates staged copies of important scenarios in closed course environments. This allows rapid repetition of rare events that would otherwise require months or years of real-world driving to encounter sufficient times for adequate training data collection.
Solution Approach 2:
The system implements continuous iterative cycles of: (1) identifying important scenarios from real-world testing, (2) staging them in closed courses, (3) collecting training data, and (4) updating models. This periodic process systematically improves model accuracy over time without requiring proportionally increasing real-world testing durations.
3Reliability
If autonomous vehicles are tested extensively in real-world conditions to improve safety, then model reliability increases, but the risk of encountering unsafe or abnormal responses from insufficiently trained models increases
Solution Approach 1:
The system performs preliminary identification and staging of challenging scenarios in controlled closed course environments before deploying models to real-world autonomous operation. This advance testing in safe conditions allows the model to learn from edge cases without the risk of unsafe responses during actual autonomous driving.
Solution Approach 2:
The system creates a protective layer of staged training scenarios that cushion against potential unsafe responses by pre-exposing the model to challenging situations in controlled environments. This preparation reduces the likelihood of abnormal or unsafe responses when the model encounters similar scenarios during real-world autonomous operation.
Data Source
AI summary
The present technology pertains to obtaining sensor data and processed sensor data related to a base world scenario encountered by an AV entity. The sensor data and processed sensor data may be assessed to determine an importance value for the base world scenario. When the importance value is above a threshold, a closed course staging system may stage a re-creation of the base world scenario in a closed course environment (i.e., a closed course scenario). An AV may then interact with the closed course scenario. Sensor data and processed sensor data from the AV's interaction with the closed course scenario may then be added to training data used to train ML models for AVs.