Autonomous Vehicle Model Update via Global Repository
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous and semi-autonomous vehicles and devices face performance failures due to adverse conditions and inconsistent sensor data, which existing systems struggle to address effectively.
Innovation Solution
A method where a task-execution device applies local-model updates based on global-model updates from a repository, generating and executing task policies, and providing feedback for global-model updates, enabling learning from other devices' experiences and facilitating collective task performance improvements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If autonomous vehicles rely on local models and sensors for task execution, then device complexity is reduced and response speed is improved, but reliability deteriorates due to adverse conditions and inconsistent sensor data
Solution Approach 1:
The system divides the model architecture into hierarchical segments: local models at individual devices and a global model at the repository. This segmentation allows each component to have specialized functionality while maintaining overall system reliability without requiring every device to handle all complexity.
Solution Approach 2:
The global model repository acts as an intermediary between individual devices and the broader system knowledge. It receives, aggregates, and processes local model updates from multiple devices, then distributes improved global models back to devices, thereby improving reliability without increasing individual device complexity.
2Adaptability or versatility
If autonomous vehicles use fixed local models for task execution, then device complexity is minimized, but adaptability deteriorates when facing adverse traffic conditions or inconsistent sensor data
Solution Approach 1:
The system implements a feedback mechanism where local models continuously send performance data and update information to the global model repository. The repository processes this feedback and generates improved global models that are distributed back to devices, enabling continuous adaptation to adverse conditions without requiring complex local update mechanisms.
Solution Approach 2:
The system merges knowledge from multiple local models into a single global model at the repository. By combining experiences and data from numerous devices, the global model achieves high adaptability to various conditions, which is then distributed back to individual devices without requiring each device to independently develop complex adaptive capabilities.
3Reliability
If autonomous vehicles implement real-time model updates based on local observations, then adaptability is improved, but loss of time increases due to continuous model training and updating
Solution Approach 1:
The system performs model updates in advance by continuously training the global model at the repository using aggregated data from multiple devices. When a device needs an update, the improved model is already prepared and can be quickly distributed, eliminating the need for time-consuming real-time training at individual devices while maintaining high reliability.
Solution Approach 2:
The system creates copies of the trained global model and distributes them to individual devices. Instead of each device independently training models in real-time, devices receive updated model copies from the repository, significantly reducing update time while maintaining task performance reliability through centralized pre-training.
Data Source
AI summary
An embodiment takes the form of a method carried out by a task-execution device. The task-execution device applies a first local-model update to a local model of the task-execution device. The first local-model update is applied based on a first global-model update to a global model of a global-model repository. The task-execution device generates an execution policy based on (i) a received task request identifying a requested task and (ii) the local model including the first local-model update. The task-execution device executes a performance of the requested task based on the generated execution policy, and obtains an observation of the performance of the requested task. Additionally, the task-execution device applies a second local-model update to the local model based on the obtained observation of the performance of the requested task, and provides the global-model repository with an indication of the second local-model update.