JapanManufacturing2024 – ongoingFujikura
Reinforcement learning for lens-alignment equipment
Fujikura reported 14.5-second lens alignment with reinforcement learning
Problem
Conventional automatic lens alignment sometimes missed product specifications and required manual correction. Training a controller directly on production equipment risked disruption and equipment wear.
Approach
Fujikura trained a reinforcement-learning controller in a simulator and transferred it to manufacturing equipment. Engineers used numerical features extracted from beam images to reduce sensitivity to visual noise. The company reported completed deployment and ongoing maintenance.
Results
Takeaway
Extracting beam features helped the controller cope with differences between simulated and real images. The timing figures came from equipment experiments, rather than a factory-wide productivity measurement.
Categories
- Size
- Large enterprise
- Industry
- Manufacturing
- Business function
- Production and quality
- Innovation type
- Process innovation
- AI technology
- Optimization
- Adoption
- Built in-house
- Evidence
- Company's own claim
Sources
Source summaries
Fujikura described simulator training and deployment of its lens-alignment controller. Equipment experiments averaged 14.5 seconds with reinforcement learning, compared with 57.9 seconds manually and 26.3 seconds with conventional automation.
Full details are in the original sources.
