Machine Learning Update Using Human-Verified Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for managing inventory in materials handling facilities face challenges in accurately tracking and confirming events, such as item removals, due to limitations in automated systems' confidence levels and the need for human verification, especially with ambiguous or low-quality sensor data.
Innovation Solution
The system distributes inquiry data to associates for processing, using a combination of sensor data and tentative values, with supplemental information like bounding boxes and velocity data, to enhance accuracy and confidence in event confirmation, and utilizes a consensus-based approach to generate output data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated systems are used to track and confirm events, then productivity is improved, but measurement precision deteriorates due to low confidence levels
Solution Approach 1:
The patent introduces an intermediary verification process where sensor data is first processed by automated systems to generate tentative values, then presented to associates for confirmation. This intermediary step allows automated processing to handle routine cases efficiently while human judgment resolves ambiguous situations, thereby maintaining both productivity and measurement precision.
Solution Approach 2:
The system implements feedback loops where associate confirmations are used to train machine learning models, improving the automated system's accuracy over time. The feedback mechanism allows the system to learn from human judgments and progressively reduce the need for manual verification while maintaining high precision.
2Measurement precision
If human verification is used to confirm events, then measurement precision is improved, but loss of time increases due to manual processing
Solution Approach 1:
The system applies partial human verification by only involving associates when automated confidence levels fall below a threshold. Routine events with high confidence are processed automatically without human intervention, while only ambiguous cases require manual verification. This partial action approach maintains measurement precision for critical events while minimizing time loss.
3Measurement precision
If more sensor data is collected to improve accuracy, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent segments the data processing function across multiple components: sensor data collection, automated tentative value generation, and human verification. This segmentation allows the system to use multiple data sources (sensors, images, audio) without requiring complex integration logic, as each component handles its specific data type independently before combining results.
Data Source
AI summary
A system may use sensor data from a facility to generate tentative values associated with an event, such as the identification of an item removed from a shelf of the facility. A confidence value associated with each of the tentative values may be less than a confidence threshold. In response, inquiry data seeking confirmation of a tentative value from an associate is generated and sent to one or more associates in the facility. Responses from the associates are collected to determine a selection of one of the tentative values. The selected tentative value is designated as output data for the system. Thereafter, the output data and the original sensor data are designated as training data, which can then be used to train or update machine learning systems. Subsequent use of the updated machine learning systems can yield more accurate results.