AGV Vision Supervision for Real-Time On-Site Neural Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated guided vehicles (AGVs) often perform poorly in real-world environments that differ significantly from the default datasets used for training, leading to low recognition rates due to variations in lighting, worker uniforms, and facility infrastructure, making it difficult for manufacturers to adapt the systems to each customer's facility.
Innovation Solution
An AGV system equipped with a mobile base, two cameras (one with a content filter for marker recognition), and a main control module that executes a recognition neural network program and a supervisor program, allowing for real-time supervised learning by updating the default dataset with user-input markers and real-world information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the AGV uses a default dataset for training the neural network, then the system can be manufactured and deployed quickly, but the recognition performance deteriorates when the actual environment differs from the default dataset
Solution Approach 1:
The system performs preliminary training with a default dataset before deployment, enabling quick manufacturing and initial operation. The preliminary action is complemented by ongoing data collection and retraining in the actual environment to adapt to specific conditions, thus maintaining both quick deployment and high recognition accuracy.
Solution Approach 2:
The neural network training process is made dynamic by allowing continuous collection of real-world image data and periodic retraining of the network. This transforms the static default dataset into a dynamically updating dataset that adapts to the specific facility environment, resolving the contradiction between quick deployment and environment-specific accuracy.
2Reliability
If the AGV manufacturer creates custom datasets for each customer facility, then the recognition performance improves, but the complexity and cost of manufacturing increases
Solution Approach 1:
The AGV system performs self-service by automatically collecting image data in the actual facility environment and using this data to retrain its own neural network. This eliminates the need for manufacturers to manually create custom datasets for each facility, reducing manufacturing complexity while maintaining high recognition accuracy through automatic adaptation.
Solution Approach 2:
The system changes the parameters of the neural network through automatic retraining using real-world data collected in the facility. This parameter adjustment process allows the system to adapt to environment-specific conditions without requiring complex manual dataset creation, thus improving recognition rate while keeping manufacturing complexity low.
3Adaptability or versatility
If the AGV operates in diverse real-world environments, then the versatility of the system improves, but the recognition performance deteriorates due to environmental variations
Solution Approach 1:
The system achieves both versatility and reliable recognition by making the neural network dynamic. It continuously collects data from diverse environments and retrains the network to adapt to each specific environment. This dynamic adaptation allows the AGV to operate in various facilities while maintaining high recognition accuracy in each location.
Solution Approach 2:
The system implements feedback by collecting real-world image data during operation and using this feedback to retrain the neural network. This closed-loop approach allows the AGV to learn from actual environmental variations and improve its recognition performance specifically for each facility it operates in, thus maintaining both versatility and reliability.
Data Source
AI summary
An automatic guided vehicle (AGV) comprising a mobile base including a drive train and configured to drive the AGV in a self-navigation mode within a facility, a first camera configured to capture first image data of objects within the facility, a second camera configured to capture second image data of objects within the facility, the second camera including a content filter, and a main control module configured to receive the first and second image data from the first and second cameras. The main control module executes a recognition neural network program. The neural network program recognizes targets in the first image data. The main control module also executes a supervisor program under user control. The supervisor program is receives the second image data and recognizes markers attached to targets in the second image data. The supervisor program produces a supervised outcome in which targets to which markers are attached are associated with categories based on user commands. The supervised outcome adjusts weights of nodes in the recognition neural program.