Autonomous Vehicle Fleet Computing for Online Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles often lack sufficient computational resources and trained models to handle unusual or unencountered scenarios, limiting their ability to operate safely and efficiently, especially when encountering conditions like sensor noise from sunlight at intersections.
Innovation Solution
Utilizing unused computational resources of autonomous vehicles to train and test machine-learning models online using live sensor data, allowing for the diversion of vehicles from service to conduct training or testing, or performing these tasks during computationally light operations, such as driving on straight roadways with low traffic, and leveraging distributed computing architectures to share resources and improve processing power.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If autonomous vehicles operate with limited computational resources and pre-trained models only, then service availability is maintained, but the ability to handle unusual or unencountered scenarios is insufficient
Solution Approach 1:
The system performs preliminary training of machine learning models during periods when computational resources are underutilized (computationally light operations). This allows the vehicle to prepare and store trained models in advance, so that when unusual scenarios are encountered, the vehicle can immediately apply pre-trained models without requiring intensive real-time computational resources.
Solution Approach 2:
The system dynamically adjusts its operational mode based on computational resource availability. During computationally light operations, the vehicle transitions to training mode to improve its models. When unusual scenarios are detected, the vehicle can divert from service to perform intensive training or testing. This dynamic adaptation allows the system to balance service availability with model improvement needs.
2Measurement precision
If autonomous vehicles divert from service to conduct training or testing, then model accuracy improves, but service performance may be compromised
Solution Approach 1:
The system performs training and testing actions partially - only diverting from service when computational resources are available and service demand is low. Rather than continuously diverting for training, the vehicle performs training during computationally light operations or when service demand allows, achieving model improvement while minimizing impact on service performance.
Solution Approach 2:
The system implements periodic training and testing cycles rather than continuous operations. The vehicle alternates between service mode and training/testing mode based on computational resource availability and service demand. This periodic approach allows the vehicle to maintain service performance while periodically improving its models through training and testing.
3Adaptability or versatility
If machine learning models are trained offline only, then computational resources are preserved during operation, but the models lack adaptability to new scenarios
Solution Approach 1:
The system performs preliminary training during offline periods and computationally light operations, preparing models in advance before they are needed. This allows the vehicle to conduct extensive training without losing operational time, as the training is completed beforehand when the vehicle is not actively serving or during low-demand periods.
Solution Approach 2:
The system extends training activities from purely offline to a continuous process that includes offline training, training during computationally light operations, and training during diverted service periods. This continuous approach ensures that model improvement is an ongoing process rather than a periodic interruption, maximizing adaptability while minimizing impact on service.
Data Source
AI summary
A method and system of using excess computational resources on autonomous vehicles. Such excess computational resources may be available during periods of low demand, or other periods of idleness (e.g., parking). Where portions of computing resources are available amongst a fleet of vehicles, such excess computing resources may be pooled as a single resource. The excess computational resources may be used, for example, to train and/or test machine-learning models. Performance metrics of such models may be determined using hardware and software on the autonomous vehicle, for example sensors. Models having performance metrics outperforming current models may be considered as validated models. Validated models may be transmitted to a remote computing system for dissemination to a fleet of vehicles.