Over the past decade, Mohamed Khan has built a career in Amazon’s fulfillment centers, helping customers ship everything from books to bookshelves. Now as his assistant general manager at a fulfillment center in Tracy, California, Mr. Khan knows what it takes to pick, sort, and pack thousands of packages every day. And Khan knows how to make sure your product arrives undamaged.
“When I joined Amazon, I started by picking products according to customer orders and had a hands-on understanding of what it takes to get the right product to the customer,” Khan said.
Khan explained that up to five employees use six visual checks to assess products for damage as they move through the fulfillment center operations. He said it’s a time-consuming task and difficult to keep in mind because employees rarely find damaged items in Amazon’s inventory.
That’s why a team of scientists at Amazon Fulfillment Technologies in Berlin, Germany are working hard to help Kahn and his colleagues. Researchers are developing advanced artificial intelligence (AI) capabilities that can spot anomalies and flag defective products before they ship.
Like other advanced AI tools, damage detection technology relies on software and large amounts of data. However, product breakage is so rare that it lacks the data needed to train the AI. The scarcity of data and the sheer size of Amazon’s diverse inventory have made AI damage detection very difficult to date. Last year, a research team determined that machine learning models could be provided with reference images to teach them how to compare images of the product they’re “seeing” with what that product should look like. To achieve this, the company uses computer vision to scan every item passing through a warehouse outside the German capital. A machine learning model then analyzes the scans to discover hidden patterns and continuously improve the system’s ability to detect damage. This approach to machine learning and computer vision will enable AI to make subjective decisions about the harm that humans do all the time.
“We looked at millions of example images of damaged and undamaged items to build our own mental model of what the item should look like,” said Jeremy Wyatt, director of applied science at Amazon Robotics. We started by training machine learning on “Then, during operation, we show an image of a specific ‘query’ item and previous ‘gallery’ images of the same product, and the AI model compares those images. This was one of several approaches we took to significantly improve performance. ”
Christoph Schwertfeger, Applied Science Manager at Amazon Fulfillment Technologies, said AI systems are three times more effective than manually identifying damaged products. Following its success, plans are being made to introduce the system to other facilities.
“Of course, the technology is only focused on this task, but our operations staff do so much every day. We plan to extend the use of this technology beyond the proving ground.” said Schwertfeger. “We hope to roll out our damage detection software to more than a dozen of our offices in North America and Europe before the holiday season. It will be able to scan for damage on your phone and will be part of the way we make sure people get undamaged gifts this holiday season.”
Schwertfeger was quick to point out that while machine learning models continue to improve as datasets grow, the true power of technology lies in the way it assists humans in their work.
“These models are convenient, but they can also be deceiving,” added Schwertfeger. “When something like that happens, we get immediate feedback from Amazon’s fulfillment network operations. Our claims experts directly teach AI how to make better decisions in the future. It is this collaboration between and machine that yields better results for our customers.”
Khan is interested in how this technology can help improve the efficiency of fulfillment centers.
“This tool has the potential to streamline tasks and help our customers manage costs and delivery times,” says Khan. “It is also very interesting because it frees up our operations employees to focus on other core tasks and activities, especially important responsibilities such as safety.”
The Amazon Fulfillment Technologies team plans to expand the capabilities of the system in addition to installing the system in more locations.
“Looking at future applications of this technology, one possibility is that we can do more than just detect damage before an order is shipped,” says Wyatt. “For example, we may be able to identify when and where the damage occurred in the first place.”
With Amazon and AI, the future will arrive better than ever.Click here for details Amazon AI.
