How Coupang Improved Warehouse Automation with Labellerr AI

As warehouse automation continues to scale across the retail and logistics industry, the ability to accurately identify and classify products becomes critical especially when dealing with perishable goods that require special handling.

Coupang partnered with Labellerr to build a high-quality image annotation pipeline capable of supporting AI models for warehouse automation. The project focused on segmenting perishable and non-perishable items inside trays and warehouse containers, even in highly cluttered and occluded environments.

Over the course of the engagement, Labellerr annotated more than 30,000 images while also performing classification and QA-based annotation tasks to improve data reliability and model readiness.

About Our Client

Coupang is one of the largest e-commerce and technology-driven retail companies in South Korea, known for its large-scale logistics infrastructure and fast delivery network.

The company operates highly automated fulfillment and warehouse systems designed to manage massive product volumes efficiently while maintaining delivery speed and inventory accuracy.

As warehouse automation continues to evolve, Coupang has been investing in AI and computer vision technologies to improve operational intelligence across inventory management, product handling, and logistics workflows.

For this project, Coupang aimed to strengthen its warehouse automation pipeline by developing AI models capable of identifying and classifying perishable items within complex tray environments, enabling smarter handling and storage decisions across fulfillment operations.

The Challenge

Modern warehouse automation systems rely heavily on computer vision models to identify products and make handling decisions in real time.

For Coupang, one of the key operational challenges was distinguishing perishable items from non-perishable products inside trays containing multiple overlapping objects.

The project presented several technical difficulties:

• Multiple items packed closely together in trays
• Heavy object occlusion
• Complex product boundaries
• Similar visual appearance between categories
• Requirement for highly accurate segmentation masks
• Need for additional QA classification workflows

Perishable goods require special warehouse handling because they are more sensitive to temperature, storage duration, and transportation conditions. Misclassification could directly impact inventory quality and operational efficiency.

To build reliable AI models for warehouse automation, Coupang required highly accurate annotated datasets capable of handling real-world warehouse complexity.

The Objective

The primary objective of the project was to create a large-scale, high-accuracy annotated dataset for training computer vision models that could:

• Identify perishable items within trays
• Distinguish them from non-perishable products
• Handle partially visible or occluded objects
• Improve automated warehouse decision-making
• Support scalable warehouse automation pipelines

In addition to segmentation, the project also included image-level classification and QA annotation tasks involving category-related validation questions for tray contents

The Solution

Labellerr designed a specialized annotation workflow tailored for warehouse inventory imagery.

The project involved semantic and instance-level segmentation of multiple products within trays containing both perishable and non-perishable items.

High-Precision Segmentation

The core requirement of the project was segmentation annotation across trays containing multiple products.

Our team annotated:

• Fruits
• Vegetables
• Packaged perishables
• Mixed grocery items
• Non-perishable packaged goods

Many of these products appeared in dense arrangements where edges overlapped or were partially hidden behind other objects.

To maintain accuracy:

• Annotators followed strict segmentation guidelines
• Occluded items were still labeled based on visible context
• Edge refinement workflows were implemented for difficult cases
• Multi-stage review processes ensured annotation consistency

This level of detail helped create a high-quality dataset suitable for training robust warehouse vision models.

Handling Occlusions and Complex Item Arrangements

One of the most technically demanding parts of the project involved segmentation of occluded objects.

In real warehouse environments, products rarely appear in perfectly isolated conditions. Items are stacked, rotated, partially covered, or tightly packed together.

For Coupang’s automation pipeline, missing a partially visible perishable item could lead to incorrect downstream decisions.

To address this, our annotation teams were trained to:

• Identify partially visible product boundaries
• Infer segmentation continuity from visible regions
• Maintain consistency across difficult edge cases
• Avoid over-segmentation in crowded trays

This helped generate datasets that better reflected real operational warehouse conditions rather than idealized training data.

Classification and QA Annotation Tasks

In addition to segmentation, the project also included classification and quality assurance annotation tasks.

Our team answered structured annotation questions related to the contents of trays and item categories. These QA annotations helped enrich the dataset with additional metadata beyond segmentation masks alone.

The classification workflow included validation tasks such as:

• Determining whether trays contained perishable items
• Categorizing product groups
• Verifying annotation consistency
• Confirming item visibility and classification accuracy

By combining segmentation with QA-driven annotation layers, the final dataset became more useful for multiple AI workflows simultaneously.

This allowed Coupang to potentially support:

• Automated sorting systems
• Inventory intelligence
• Smart warehouse routing
• Product handling prioritization
• Perishable goods monitoring

Scale of the Project

The engagement involved annotation of more than 30,000 images across multiple warehouse scenarios and tray configurations.

Managing annotation quality at this scale required:

• Structured annotation guidelines
• Dedicated QA review cycles
• Consistency validation across teams
• Continuous edge-case monitoring
• Iterative feedback loops

The large dataset volume also meant maintaining consistency across thousands of complex segmentation tasks while preserving high annotation precision.

Conclusion

As logistics and fulfillment operations continue adopting AI-driven automation, the need for highly accurate training data becomes increasingly important.

Our collaboration with Coupang focused on solving a real-world warehouse intelligence challenge through large-scale segmentation and classification annotation.

By annotating over 30,000 images with high precision, including difficult occluded and overlapping objects, we helped create datasets tailored for warehouse automation systems that depend on reliable identification of perishable goods.

The project demonstrates how high-quality annotation workflows can directly support the development of more intelligent, efficient, and scalable AI-powered logistics operations.

‍