Machine Learning Perception Fields for Unknown Driving Scenarios
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous driving (AV) and advanced driving assistance systems (ADAS) face limitations in accuracy due to reliance on datasets from recorded situations, which can restrict their ability to respond effectively to unknown scenarios.
Innovation Solution
The implementation of perception fields, a learned representation of road objects as virtual force fields, allows the system to sense and respond to road objects through a mathematical function dependent on spatial position, enabling the vehicle to handle unknown situations by decomposing actions into fundamental components and providing a holistic approach for driving policies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the machine learning process is trained solely on recorded situations, then the training data is available and easy to obtain, but the accuracy and ability to handle unknown scenarios is limited
Solution Approach 1:
The patent applies preliminary action by training the machine learning process in advance on simulated edge cases and unknown scenarios before deployment. This pre-training on diverse synthetic data prepares the system to handle situations that may not be present in recorded real-world data, thereby improving both accuracy and adaptability to unknown scenarios.
Solution Approach 2:
The patent uses copying by creating synthetic training data that replicates edge cases and unknown scenarios through simulation. These copied virtual scenarios are then used to train the machine learning model, enabling it to learn from a broader range of situations without requiring extensive real-world data collection for every possible scenario.
2Adaptability or versatility
If more diverse training scenarios are used to improve adaptability, then the ability to handle unknown scenarios improves, but the complexity of data collection and processing increases
Solution Approach 1:
The patent replaces the mechanical system of physical data collection with a computational simulation system. Instead of physically capturing diverse scenarios in the real world, the system uses virtual simulations to generate training data, significantly reducing the complexity of data collection and processing while maintaining high adaptability.
Solution Approach 2:
The patent applies parameter changes by systematically varying parameters in simulated environments to generate diverse training scenarios. By changing environmental conditions, object positions, and scenario parameters in simulations, the system efficiently creates comprehensive training data without the logistical complexity of real-world data collection.
Data Source
AI summary
A method for training and testing a machine learning process, the method includes (a) learning virtual fields based on simulations of behaviors of a vehicle when faced with situations involving objects within environments of the vehicle, the virtual fields represent potential impacts of objects on the behaviors of the vehicle, wherein the learning is based on a virtual physical mode; (b) training the machine learning process to generate the virtual fields by applying a training process that uses outcomes of the simulations to provide a trained machine learning process; and (c) testing the trained machine learning process by feeding the trained machine learning process with other situations to provide test results.