Autonomous Driving Corner-Case Testing with Human Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for testing intelligent driving vehicles are time-consuming and costly, with conventional tests being unsuitable for safety verification due to the high fitting and low generalization of intelligence algorithms, and reinforcement learning methods face low testing efficiency and potential omission of dangerous corner cases.
Innovation Solution
An intelligent driving test method that combines active exploration by an original driver with reproduction by an imitation learning driver, utilizing a dynamic corner case completion library based on human feedback to enhance testing efficiency and completeness, where the original driver explores and the imitation learning driver reproduces corner cases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If reinforcement learning is used to build an original scene driver for exploration, then the ability to discover new corner cases is improved, but the testing efficiency deteriorates due to low exposure rate of new corner cases
Solution Approach 1:
The patent creates a copy of the original scene driver's behavior through behavior correction and imitation learning. The imitation learning driver replicates the exploration capabilities of the original driver while operating more efficiently, effectively copying the corner-case-discovery function without the efficiency penalties of the original reinforcement learning approach
Solution Approach 2:
The patent introduces behavior correction as an intermediary mechanism between the original scene driver and the testing environment. This intermediary layer filters and refines the exploration behavior, allowing the system to maintain the ability to discover new corner cases while improving testing efficiency through more targeted exploration
2Reliability
If manual setting of test scenes is performed to achieve high coverage, then the completeness of testing is improved, but the cost and difficulty increase significantly
Solution Approach 1:
The system performs self-service by automatically generating and optimizing test scenes through the original scene driver's exploration and the imitation learning driver's reproduction. The behavior correction mechanism automatically identifies and focuses on critical corner cases without human intervention, enabling the system to achieve high testing completeness autonomously
Solution Approach 2:
The patent performs preliminary action by pre-training the imitation learning driver on corner cases discovered during the behavior correction phase. This preliminary learning enables the system to quickly reproduce and test critical scenarios without requiring manual setup for each test case, significantly reducing the difficulty and cost of achieving comprehensive coverage
3Adaptability or versatility
If the original scene driver explores continuously to adapt to personalized intelligence algorithms, then the adaptability is improved, but the testing efficiency deteriorates due to low sample utilization
Solution Approach 1:
The imitation learning driver creates an efficient copy of the exploration process, capturing the adaptability lessons learned by the original driver and reproducing them with higher sample utilization. This copying mechanism allows the system to maintain adaptability to personalized algorithms while dramatically improving testing efficiency
Solution Approach 2:
The behavior correction mechanism implements feedback by continuously monitoring the original scene driver's exploration and using human feedback to refine the behavior. This feedback loop enables the system to adapt to personalized intelligence algorithms while efficiently identifying and focusing on the most critical corner cases for testing
Data Source
AI summary
Disclosed is an intelligent driving test method with corner cases dynamically completed based on human feedback, including the following steps: obtaining an initial state of a real environment; building an original scene driver based on reinforcement learning, and correcting behavior selection to obtain an exploratory behavior; testing in a testing environment, performing expert evaluation, and building a dynamic corner case completion library based on the human feedback; building an imitation learning driver based on the human feedback, updating a policy based on test data in the dynamic corner case completion library, training the imitation learning driver, and outputting a corner case reproduction behavior; obtaining the initial state of the real environment and an initial environmental state of the dynamic corner case completion library, and selecting a scene driver; and outputting a corresponding behavior, and testing in the testing environment based on the corresponding behavior to obtain a test result.