AV Cost Function Training Using Manual Driving Deviations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately interpreting and responding to their surroundings due to imbalances in cost functions, requiring continuous adjustments to navigate complex environments and mimic human driving behavior.
Innovation Solution
Training cost and value functions using machine learning models and algorithms, such as decision trees and neural networks, by comparing manual driving data with autonomous vehicle predictions to identify and update cost functions based on deviations, utilizing manual driving data to refine AV control systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If cost functions are continuously adjusted to mimic human driving behavior, then navigation accuracy is improved, but system complexity and resource consumption increase
Solution Approach 1:
The system performs preliminary actions by collecting and storing manual driving data in advance, creating a comprehensive dataset that captures human driving behavior patterns. This pre-collected data serves as training material for machine learning models, allowing the system to learn optimal cost function parameters without requiring continuous real-time adjustments, thereby improving navigation accuracy while managing system complexity
Solution Approach 2:
The system creates copies of human driving behavior through machine learning models that are trained on recorded manual driving data. These models generate synthetic training instances that replicate human decision-making patterns, allowing the autonomous vehicle system to learn from multiple virtual examples without requiring continuous physical testing, thus improving accuracy while reducing the complexity of real-world experimentation
2Reliability
If manual driving data is collected and processed to train cost functions, then behavior alignment with human drivers is improved, but data processing time and computational resources increase
Solution Approach 1:
The system segments the data processing workflow into distinct phases: data collection during manual driving, offline processing to generate training instances, and model training. By dividing the comprehensive manual driving dataset into smaller, manageable training instances that capture specific behavioral patterns, the system can process data more efficiently while maintaining reliable behavior alignment through targeted learning of specific driving scenarios
Solution Approach 2:
The system performs preliminary processing of manual driving data to generate training instances before actual model training begins. This pre-processing stage creates structured training data with labeled features and outcomes, allowing the machine learning models to learn efficiently from prepared material rather than raw data, thereby reducing computational time and resources while improving behavior alignment
3Measurement precision
If deviations between manual driving and autonomous predictions are analyzed, then cost function accuracy is improved, but computational overhead increases
Solution Approach 1:
The system applies partial action by analyzing only the most significant deviations between manual driving data and autonomous predictions, rather than processing every single data point. By identifying and focusing on training instances where the autonomous system's predictions diverge meaningfully from human behavior, the system improves cost function accuracy through targeted analysis of critical scenarios while reducing overall computational overhead through selective processing
Data Source
AI summary
Techniques are disclosed for training one or more cost functions of an autonomous vehicle (“AV”) control system based on difference between data generated using the AV control system and manual driving data. In many implementations, manual driving data captures action(s) of a vehicle controlled by a manual driver. Additionally or alternatively, multiple AV control systems can be evaluated by comparing deviations for each AV control system, where the deviations are determined using the same set of manual driving data.