Driving Scenario Sampling With Simulated Driver Behavior

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Autonomous vehicles require high-quality training data that captures diverse driving scenarios to ensure accurate machine learning model performance, but real-world data collection is limited and lacks comprehensive representation of driver behaviors.

Innovation Solution

A method involving the simulation of driving scenarios using randomly initialized physical and mental states of objects, such as virtual vehicles and pedestrians, to generate diverse and comprehensive training data, which includes assigning initial positions, accelerations, and mental states like acceleration preferences and politeness factors, and excluding failed scenarios like collisions to train machine learning models like deep neural networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If real-world data collection is used to train machine learning models, then the training data reflects actual driving conditions, but the data lacks comprehensive representation of diverse driver behaviors and scenarios

Engineering Contradiction:

Improverepresentation of driver behaviorsVSAvoiddiversity of driving scenarios

Core Design Contradiction:

Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent creates virtual copies of driving scenarios through simulation rather than relying solely on real-world data collection. Virtual vehicles with simulated driver behaviors replicate real driving conditions while allowing comprehensive control over scenario diversity. This copying approach enables training on a much broader range of situations including rare and edge cases that would be difficult to capture in real-world data.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The system varies multiple parameters simultaneously to generate diverse driving scenarios: driver behavior parameters (aggression, politeness, acceleration preferences), environmental parameters (weather, time of day, location), and scenario parameters (traffic density, vehicle types). By systematically changing these parameters, the simulation generates comprehensive training data covering the full spectrum of possible driving conditions.

Inventive Principle:

Principle #35Parameter changes

2Adaptability or versatility

If simulated driving scenarios are generated with random initial states, then diverse driving scenarios can be created, but the quality and realism of the training data may be compromised

Engineering Contradiction:

Improvediversity of driving scenariosVSAvoidquality of training data

Core Design Contradiction:

Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The system performs preliminary actions by carefully designing and configuring virtual vehicles with realistic physical properties, sensor models, and driver behavior parameters before generating scenarios. The simulation environment is pre-configured with accurate road networks, traffic rules, and environmental conditions. This preliminary preparation ensures that randomly generated scenarios maintain high realism and quality throughout the training process.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The simulation system incorporates feedback mechanisms to monitor and evaluate the quality of generated scenarios. Failed scenarios (those with collisions or unrealistic behaviors) are identified and excluded from training data. The system uses feedback from scenario outcomes to refine and adjust simulation parameters, ensuring continuous improvement of training data quality while maintaining diversity.

Inventive Principle:

Principle #23Feedback

3Quantity of substance

If all simulated driving scenarios are used for training, then the training dataset is comprehensive, but failed scenarios with collisions reduce model accuracy

Engineering Contradiction:

Improvesize of training datasetVSAvoidaccuracy of machine learning model

Core Design Contradiction:

Quantity of substanceVSReliability

Solution Approach 1:

The system extracts and separates failed scenarios (those resulting in collisions or unrealistic outcomes) from the complete set of simulated driving scenarios. These failed scenarios are explicitly excluded from the training dataset, while successful scenarios are retained. This extraction process ensures that the training data maintains high quality and reliability by removing problematic examples that would degrade model performance.

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentUS11938957B2Driving scenario sampling for training/tuning machine learning models for vehicles

Publication Date: 2024.03.26 MOTIONAL AD LLC

AI summary

Enclosed are embodiments for sampling driving scenarios for training machine learning models. In an embodiment, a method comprises: assigning, using at least one processor, a set of initial physical states to a set of objects in a map for a set of simulated driving scenarios, wherein the set of initial physical states are assigned according to one or more outputs of a random number generator; generating, using the at least one processor, the set of simulated driving scenarios in the map using the initial physical states of the objects in the set of objects; selecting, using the at least one processor, samples of the simulated driving scenarios; training, using the at least one processor, a machine learning model using the selected samples; and operating, using a control circuit, a vehicle in an environment using the trained machine learning model.