USPS Turns to AI at the Edge to Boost Package Processing

An edge artificial intelligence platform recently implemented by the U.S. Postal Service is helping federal insiders trace the many millions of packages it processes each week in a matter of hours instead of multiple days.

Open-source software from Nvidia is also delivering AI models to make more use of data collected in systems at 195 Postal Service sites across the nation, the company confirmed Thursday.

“The federal government has been for the last several years talking about the importance of artificial intelligence as a strategic imperative to our nation, and as an important funding priority. It's been talked about in the White House, on Capitol Hill, in the Pentagon. It's been funded by billions of dollars and it's full of proof of concepts and pilots,” Nvidia’s Vice President of Federal Anthony Robbins told Nextgov Tuesday. “And this is one of the few enterprisewide examples of an artificial intelligence deployment that I think can serve to inspire the whole of the federal government.”

The Postal Service hasn’t generally been associated with speedy deliveries, and the pandemic spurred major delays for “snail mail” over the last year. Still, officials said it’s now tapping AI-driven techniques to complete its mission quicker and more effectively. In separate exchanges this week, Robbins and a Postal Service spokesperson briefed Nextgov on this tech-centered deployment and what might follow.

Roughly 129 billion pieces of mail and 7.3 billion packages were processed by the Postal Service last year and within its infrastructure are thousands of scanners, cameras and other elements. In 2019, a federal data scientist was struck with an idea to place edge AI servers in Postal Service processing centers’ systems in an effort to gain and share more data points and insights from the billions of images that were generated as items zipped through to their destinations. 

In a three-week sprint, that official and Nvidia architects thought through models and capabilities and worked to figure out what they could do to improve processing with equipment and images of postage, damaged barcodes, package sizes and weights and other variables. A result of that engagement was the Edge Computing Infrastructure Program, or ECIP, a distributed edge AI system now running at USPS locations, via the NVIDIA EGX platform.

Todd Schimmel, the manager who oversees Postal Service systems including ECIP, said in a statement that while it used to take eight or 10 people several days to track down items, it now takes one or two just a couple hours. 

Robbins confirmed Accenture Federal Services, Dell Technologies and Hewlett-Packard Enterprise also contributed to this broader effort. Necessary gear and servers were deployed, and models were trained to complete AI tasks involving computer vision and more. Computers were added at the processing and distribution centers used by the Postal Service, its spokesperson also explained. And there were specialized computing cabinets—or nodes—that contain hardware and software specifically tuned for creating and training ML models, installed at two data centers.  

“The software being used is common to commercial implantations of current AI/ML best practices,” the spokesperson said.

The agency went from proof of concept to full-scale deployment in a year’s time, Robbins noted. He pointed to one analysis that revealed a computer task needing two weeks on a network of servers with 800 central processing units can now be completed in 20 minutes on the four Nvidia graphics processing units in the HPE server. Further, each edge server processes 20 terabytes of images a day from more than 1,000 mail sorting machines.

“The AI work that has to happen across the federal government is a giant team sport,” Robbins noted. “And the Postal Service’s deployment of AI across their enterprise exhibited just that.”

Referred to as a “digital mailperson,” Nvidia's open-source software, the Triton Inference Server, is what can help deliver AI models based on the 195 postal systems’ makeup and needs. Almost three dozen applications have already been thought up that could be deployed on top of ECIP to help further augment the Postal Service’s processing of the multitudes of pieces of mail that flow through it daily, going forward. 

One of those that might be implemented as soon as this summer, Nvidia’s release noted, could rapidly decipher a damaged barcode.

“When we started this project, they almost couldn't even imagine that what we have now done was even possible. So now, it's kind of opened their eyes,” Robbins said. “And there's people all across the Postal Service and their contractors looking at the business of processing 40% of the world's mail—and how artificial intelligence and computer vision might assist.”

The USPS spokesperson didn’t want to speculate on the potential use cases that have not been fully vetted by the organization but noted one other example may be calculating short paid postage on packages.  

“The implementation of that model would be dependent on a thorough review and benefit evaluation,” the official said.

They added that lessons are being learned every day from ECIP. One big takeaway, the spokesperson said, was that “implementation takes highly specialized resources.” Having a good architecture foundation that supports MLOps—or the practice to deploy and maintain machine learning-centered systems—is also key.

Drawing on his own experiences working with the federal government for more than three decades, Robbins emphasized how impressed he was with how quickly this AI-centered project rolled out. 

“We brought artificial intelligence and machine learning to a computer vision-, image processing-like application across the enterprise—and that same kind of work exists [in various other agencies],” Robbins said. “So there's a good model here.”