Dynamics-Dependent Behavioral Planning for Self-Driving Vehicles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Self-driving vehicles face challenges in adapting driving maneuvers due to differences in dynamics between test vehicles driven by humans and actual automated vehicles, as well as changes in vehicle conditions and weather, which existing methods struggle to address effectively.
Innovation Solution
A method for training a behavior planner using observation data from test drives, incorporating a dynamics model to translate the data to the conditions of the target vehicle, allowing for adaptation to various scenarios and vehicle changes without the need for additional test data, and utilizing a classifier model for maneuver identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If observation data from test drives in human-driven test vehicles is used to train the behavior planner, then the training process is simplified and less extensive testing is needed, but the dynamics differences between test vehicles and target automated vehicles cause inaccuracies in maneuver timing and execution
Solution Approach 1:
A dynamics model is introduced as an intermediary component that receives observation data from human-driven test vehicles and transforms it into corrected observation data that accounts for the dynamics differences of the automated target vehicle. This mediator enables accurate training without requiring physical test drives of the target vehicle, resolving the contradiction between ease of training and maneuver timing accuracy.
2Productivity
If the behavior planner is trained using raw observation data without dynamics correction, then the training process is faster and simpler, but the vehicle behavior may be surprising or inappropriate to other traffic due to dynamics mismatches
Solution Approach 1:
The dynamics model performs preliminary correction of observation data before the behavior planner training process begins. By pre-adjusting the training data to account for dynamics differences, the system ensures that the behavior planner learns appropriate maneuvers for the automated vehicle's actual capabilities, maintaining reliability while preserving training speed.
3Measurement precision
If extensive test drives in the actual target vehicle are conducted to capture accurate dynamics, then maneuver execution accuracy is improved, but the time and resources required for testing increase significantly
Solution Approach 1:
Instead of conducting extensive physical test drives in the target automated vehicle, the system creates a virtual copy of the testing process by using a dynamics model to simulate how the target vehicle would respond to observation data. This copying approach achieves accurate maneuver execution training without the time cost of extensive physical testing.
4Adaptability or versatility
If the behavior planner adapts to different vehicle conditions and weather changes, then the adaptability to diverse scenarios is improved, but the complexity of the training process and dynamics modeling increases
Solution Approach 1:
The dynamics model is designed as a universal framework that can handle multiple vehicle conditions and weather scenarios through a single unified model structure. Rather than creating separate models for each scenario, the universal dynamics model adjusts its parameters to account for different conditions, maintaining adaptability while avoiding the complexity of multiple specialized models.
Data Source
AI summary
A method for training a behavior planner for an at least partially self-driving target vehicle on the basis of observation data regarding kinematics and/or dynamics that have been recorded during at least one test drive in a test vehicle includes identifying a driving maneuver that moves the test vehicle from an initial state to an end state using the observation data, ascertaining the maneuver end time, retrieving a maneuver duration required by the target vehicle to perform the identified driving maneuver from a dynamics model of the target vehicle, labeling observation data from a time interval, defined by the maneuver duration, with the identified driving maneuver, and training the behavior planner, using the labeled observation data, to map observation data that indicate a state of the target vehicle to at least one driving maneuver to be performed.