Cross-Customer ML Model Bootstrap for Logistics Automation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing logistics systems face challenges in quickly training and testing machine-learning (ML) models for optimizing the operation of forklifts and Autonomous Mobile Robots (AMRs) in new warehouses, often requiring large datasets and resulting in less efficient operations during the data accumulation phase.
Innovation Solution
A centralized system that provides compute and storage resources for training and testing ML models, leveraging datasets from multiple customers to select the best initial ML model for new warehouses, and implementing model drift monitoring and Ideal World dataset resetting to ensure optimal performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If ML models are trained using traditional methods with large datasets from a single warehouse, then model accuracy is improved, but the time required for deployment is increased and operational efficiency deteriorates during data accumulation
Solution Approach 1:
The patent combines datasets from multiple warehouses (different customers) to train a single ML model. This merging approach allows the model to learn from diverse operational patterns across multiple environments, improving generalization and accuracy while reducing the time needed to accumulate sufficient training data in any single warehouse.
Solution Approach 2:
The trained ML model is designed to be universally applicable across multiple warehouses and different types of edge devices (forklifts, AMRs). This multi-functional model can be deployed to various customers without requiring separate training for each, significantly reducing deployment time while maintaining reliability through cross-validation across diverse datasets.
2Reliability
If data is accumulated from forklifts and AMRs before training ML models, then model reliability is improved, but operational efficiency worsens during the data accumulation phase
Solution Approach 1:
The system performs preliminary data collection from multiple warehouses simultaneously before model training begins. By pre-accumulating diverse datasets from multiple sources in parallel, the system ensures sufficient training data is available to achieve reliable models without delaying operations in any single warehouse, thus maintaining productivity during the data gathering phase.
Solution Approach 2:
Datasets from multiple warehouses are merged to create a comprehensive training corpus. This combination allows the system to achieve reliable model training faster than single-warehouse accumulation would permit, as the combined data volume and diversity are reached more quickly, thereby maintaining operational efficiency while ensuring model reliability.
3Adaptability or versatility
If ML models are customized for each individual warehouse, then adaptability is improved, but device complexity and deployment difficulty increase
Solution Approach 1:
The patent implements local quality by allowing the centralized ML model to adapt to specific warehouse characteristics through fine-tuning or parameter adjustment based on local data patterns. While the base model remains universal, local adaptations can be applied selectively to capture warehouse-specific nuances without requiring complete customization, thus maintaining manageability while achieving adaptability.
Solution Approach 2:
The system creates copies of the trained ML model for deployment to different warehouses. These model copies can be independently managed and updated from the central training system, reducing complexity by standardizing the deployment process while still allowing for local adaptations through parameter adjustments or fine-tuning of the copied models.
4Reliability
If extensive data collection is performed in new warehouses, then model accuracy is improved, but the time to deployment is increased
Solution Approach 1:
The system performs preliminary data collection and model training using aggregated datasets from multiple warehouses before deployment to new locations. This advance preparation ensures that models are pre-trained with sufficient accuracy, eliminating or minimizing the need for extensive on-site data collection at new warehouses and thereby reducing deployment time while maintaining model accuracy.
Solution Approach 2:
Data from multiple existing warehouses is merged to create a robust training dataset that provides sufficient statistical power for accurate model training. This merged dataset approach eliminates the need for lengthy data collection periods at individual new warehouses, as the combined historical data from multiple sources provides the necessary training volume and diversity upfront, thus improving accuracy without extending deployment time.
Data Source
AI summary
One example method includes determining a system aggregate drift metric score based on aggregate drift metric scores received from near-edge nodes associated with a central node. The system aggregate drift metric score indicates a level of model drift across a logistics system. The system aggregate drift metric score is compared with a system drift threshold. Current datasets are received from the near-edge nodes when a system aggregate drift metric score is greater than the system drift threshold. A first dataset is generated comprising a joining of the current datasets received from the plurality of near-edge nodes. A second dataset is received from each the near-edge nodes. The first and second datasets are used to select a machine-learning (ML) model to deploy at each of the near-edge nodes.