Autonomy Training Data Densification for Rare Safety-Critical Events
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing approaches for training autonomous vehicles (AVs) face challenges in improving safety performance due to the rarity of safety-critical events, leading to the 'Curse of Rarity' (CoR), which results in performance stagnation and biased learning from failure scenarios, known as the 'seesaw effect', hindering the development of safe AVs.
Innovation Solution
A dense learning framework is implemented using data densification techniques, including safety-critical episode selection, episode state modification, and retrospective counterfactual simulation to enhance training data, focusing on both avoidable crash and near-miss events, and utilizing an AI-based safety metric for real-time evaluation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If extensive data collection from public roads is used for training, then the amount of training data increases, but the frequency of safety-critical events remains too low to achieve required safety performance
Solution Approach 1:
The patent extracts and isolates safety-critical events from the extensive training data using automated identification algorithms. These rare events are separated from normal driving data to create a focused training subset, allowing the system to learn from safety-critical scenarios without being diluted by the vast amount of non-critical data.
Solution Approach 2:
The patent changes the parameter of data distribution by synthesizing additional safety-critical scenarios through simulation and counterfactual reasoning. This transforms the training data from naturally occurring rare events to a denser distribution of safety-critical examples, enabling effective learning despite the original rarity of such events.
2Productivity
If training focuses on safety-critical events to improve safety performance, then learning efficiency improves, but biased learning from failure scenarios occurs causing the seesaw effect
Solution Approach 1:
The patent segments safety-critical events into different categories: actual failures, near-misses, and counterfactual scenarios. This segmentation allows the system to learn from diverse safety-related experiences without overfitting to failure patterns, maintaining balanced learning across different safety scenarios.
Solution Approach 2:
The patent applies counterfactual reasoning to invert the learning approach by synthesizing scenarios where the AV made correct safety decisions. This inversion balances the bias toward failure learning by providing equally weighted examples of successful safety-critical decisions, preventing the seesaw effect.
3Adaptability or versatility
If deep learning techniques are used to address high dimensionality, then the ability to handle complex variables improves, but estimation variance of policy gradient increases due to rarity of events
Solution Approach 1:
The patent performs preliminary action by pre-identifying and pre-processing safety-critical events before the main training process. This preliminary segmentation and synthesis of critical data ensures that when deep learning models process high-dimensional variables, they work with a densified dataset that reduces estimation variance from the outset.
Solution Approach 2:
The patent introduces an intermediary data densification layer between data collection and deep learning training. This intermediary process synthesizes and densifies safety-critical events, creating a bridge that allows deep learning techniques to effectively handle high-dimensional variables without suffering from the high estimation variance caused by event rarity.
Data Source
AI summary
A system and method of training a safety-critical autonomous agent. The method carried out by the system includes: obtaining initial training data to be used for training a safety-critical autonomous agent; densifying the initial training data using a data densification process in order to generate densified training data; and training a safety-critical autonomous agent using the densified training data. The data densification process includes: selecting safety-critical episodes from the initial training data in order to generate safety-critical episode training data; and/or generating artificial safety-critical episode data representing one or more artificial safety-critical episodes, wherein the artificial safety-critical episode data is generated based on at least one safety-critical episode. The densified training data is or is based on one or both of the safety-critical episode training data and the artificial safety-critical episode data.