Simulation-Based Driving Data Collection for Human-Like AV Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training autonomous driving systems, such as reinforcement learning and imitation learning, often result in vehicle behaviors that are not human-like, leading to potentially dangerous situations on the road, and require extensive and costly data collection from real-world tests.
Innovation Solution
A method involving a simulation game where users control a vehicle agent in predefined road scenarios, generating human demonstrations that are collected and used to train a planning module for autonomous driving systems, allowing for the creation of human-like training data without the need for extensive real-world testing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If reinforcement learning is used to train the behavior planning module in a simulation tool, then the agent can explore and learn policies through trial and error, but the resulting vehicle behaviors are not human-like and may lead to dangerous situations on the road
Solution Approach 1:
The patent uses imitation learning to copy human driving behaviors by collecting steering actions from users playing a simulation game. Instead of generating behaviors through reinforcement learning exploration, the system directly copies human demonstrations from the game, ensuring human-like behavior patterns while maintaining training efficiency through simulated data collection.
2Reliability
If imitation learning is used to train the behavior planning module, then human-like behaviors can be achieved, but a large amount of training data must be collected from a large scale fleet of testing vehicles
Solution Approach 1:
The patent replaces the need for large-scale real vehicle fleets with a simulation game running on user devices. Human driving behaviors are copied through the game interface, where users control a vehicle agent in predefined scenarios. This approach maintains behavioral realism while dramatically simplifying the data collection infrastructure.
Solution Approach 2:
The simulation game acts as an intermediary between human users and the training data collection process. Instead of directly instrumenting real vehicles, the system uses the game as a mediator to capture human driving decisions in a controlled, scalable environment that can be distributed to many users without requiring physical testing vehicles.
3Quantity of substance
If a large scale fleet of testing vehicles is deployed to collect training data, then sufficient data can be gathered for imitation learning, but the process becomes costly and time-consuming
Solution Approach 1:
The patent generates training data by copying human driving behaviors through a simulation game rather than collecting data from real vehicles. Users play predefined road scenarios in the game, and their steering actions are recorded as training data. This approach produces sufficient training data volume without the time and cost constraints of real-world fleet deployment.
Solution Approach 2:
The patent prepares training data in advance through simulated gameplay before actual training deployment. By having users complete scenarios in the simulation game beforehand, the system accumulates a ready-to-use dataset that eliminates the need for time-consuming real-world data collection campaigns when training is needed.
Data Source
AI summary
The present disclosure enables a collection of training data suitable for training an autonomous driving system of a vehicle. A database is provided that stores predefined road scenarios, and user devices are provided with a simulation game for controlling a vehicle agent in a road scenario. A plurality of user devices, running the simulation game, play a road scenario from the database to control a vehicle agent in the simulation game with steering actions entered by individual users, which generates a human demonstration of each road scenario that is played. Several human demonstrations played on the plurality of user devices can be maintained by a demonstration database and made available as training data suitable for training an autonomous driving system of a vehicle. A large amount of training data can be automatically generated, without requiring expensive or time-consuming real-world tests to generate suitable training data manually.