Amazon builds AI model to optimize packaging

Amazon is leveraging artificial intelligence (AI) as part of its strategy to minimize the use of extra packaging materials.

The Package Decision Engine, an AI model that Amazon designed and built, is able to determine the most efficient type of packaging for each item it learns about. Amazon then uses this information to help reduce the number of cardboard boxes, air pillows, tape and mailers used to send purchases to customers. 

Along with other packaging innovations Amazon has been putting in place since 2015, Amazon says the engine has helped it avoid more than 2 millions tons of packaging material worldwide in the past nine years.

Built on the Amazon Web Services (AWS) cloud platform, the multimodal AI model can predict when a more durable product like a blanket doesn’t need protective packaging, or when a potentially fragile item like a set of dinner plates might need a studier box. 

It uses a combination of deep machine learning, natural language processing and computer vision, and is continuously learning about Amazon’s evolving packaging options. Its decisions are empirically accurate, according to Amazon scientists, meaning that it predicts the most efficient package choice the majority of the time.

The process includes multiple steps to gather information about each item. When an item first arrives at the Amazon fulfillment center, it is photographed in a computer vision tunnel that determines the product’s dimensions, spots defects and captures multiple images of the product. 

This also allows the model to detect if there is a bag or box around an item, or detect the presence of exposed glass. The engine also uses natural language processing and leverages text-based data from each item, such as the item’s name, description, price and package dimensions. 

It also collects information in near-real time from customer feedback that is reported through Amazon’s online returns center, product reviews and other customer feedback channels.

After compiling the information, the model produces a score that predicts the best packaging type to use. The packaging selection is remembered by the model and used to understand future packaging needs.

Amazon scientists have trained the AI model by showing it millions of examples of products that had been successfully delivered in various types of packaging without damage. They also showed it products that have arrived damaged, along with the keywords and packaging types used in each scenario.

As a result, the model has learned that certain keywords are important when making packaging decisions. For example, a padded mailer with limited cushioning might not adequately protect an item with the words "grocery," "screen," or "stoneware" in the description, so the model would recommend a sturdier option, such as a box. 

The model also learned that keywords like "multipack," "bag," "shrink" and "pack" were also associated with lower damage rates in the mailer, and indicated the product might already have protective packaging and not need additional protection.

"We wanted the ability to quickly identify the most efficient packaging option for each item, while also predicting how safely each product would ship," said Kayla Fenton, senior manager of technology products with Amazon’s Packaging Innovation team, in a corporate blog post. "The use of AI through the Package Decision Engine has allowed us to advance our packaging efficiency work at scale quickly, and it has worked so well that we’re implementing this technology across Amazon’s broader global footprint."