Autonomous Vehicle On-Board Data Selection for ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately predicting object behavior due to sparse and difficult-to-isolate real-world vehicle events in vast datasets, coupled with limited computing power and storage at the edge, and high latency and cost in cloud processing.
Innovation Solution
The system selectively transmits useful training data logs from autonomous vehicles to a remote computing device for machine learning model training, using criteria such as spatial, temporal, and map-based features to identify relevant data, thereby optimizing data usage and improving model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all sensor data logs are transmitted to cloud for training, then model accuracy is improved, but communication bandwidth costs and latency increase
Solution Approach 1:
The system extracts only the most useful and informative sensor data logs from the vehicle's sensor data, identifying and transmitting only those logs that contain valuable training information to the cloud, rather than transmitting all collected data. This extraction process filters out redundant and less useful data, reducing communication bandwidth costs while maintaining model training effectiveness.
Solution Approach 2:
The system segments the sensor data logs into different categories based on their usefulness for training, using criteria such as spatial features, temporal features, bounding box inconsistencies, and map-based features. This segmentation allows the system to selectively transmit only the most valuable segments (data logs) to the cloud, optimizing the balance between model accuracy and communication costs.
2Productivity
If more computing power is provided at edge, then data processing capability is improved, but device complexity and cost increase
Solution Approach 1:
The system performs preliminary processing and analysis of sensor data logs at the edge device (vehicle), evaluating their usefulness before transmission. By conducting initial filtering and selection operations at the edge using relatively simple criteria, the system reduces the need for complex computing power at the edge while still achieving effective data preparation for cloud-based model training.
3Loss of time
If real-world data is collected and processed at edge, then data freshness is improved, but storage and computing constraints are exceeded
Solution Approach 1:
The system extracts only the most useful and informative sensor data logs from the vehicle's sensor data, identifying and transmitting only those logs that contain valuable training information to the cloud, rather than transmitting all collected data. This extraction process filters out redundant and less useful data, reducing communication bandwidth costs while maintaining model training effectiveness.
Solution Approach 2:
The system performs preliminary processing and analysis of sensor data logs at the edge device (vehicle), evaluating their usefulness before transmission. By conducting initial filtering and selection operations at the edge using relatively simple criteria, the system reduces the need for complex computing power at the edge while still achieving effective data preparation for cloud-based model training.
Data Source
AI summary
Systems and methods for on-board selection of data logs for training a machine learning model. The methods include, by an autonomous vehicle, receiving sensor data logs corresponding to surroundings of the autonomous vehicle from a plurality of sensors, identifying one or more events within each sensor data log. The methods also include, for each sensor data log: analyzing features of the identified one or more events within that sensor data log for determining whether that sensor data log satisfies one or more usefulness criteria for training a machine learning model, and transmitting that sensor data log to a remote computing device for training the machine learning model if that sensor data log satisfies one or more usefulness criteria for training the machine learning model. The features can include spatial features, temporal features, bounding box inconsistencies, or map-based features.