Army using ‘sprints’ to find stumbling blocks for AI in acquisition programs

The Army says it’s in the midst of 100 days of “sprints” to help figure out how to incorporate artificial intelligence into its acquisition programs. The objectives are to let the Army develop its own AI algorithms in a process officials feel confident they can trust, but also to bring in third-party ones in a secure way.

There are wide-ranging AI adoption initiatives throughout the Defense Department focused on things like identifying the most compelling use cases for AI and ensuring the department adopts the technology responsibly. But for the Army’s acquisition community, a big part of the current 100-day implementation plan is identifying the factors that might get in the way of real-world AI adoption, and resolving those.

“We’re aligning with everybody to look at the table stakes stuff, but we’re also looking at the obstacles that are going to make adoption of AI a little bit harder for the Army or the DoD at large,” Young Bang, the principal deputy assistant secretary of the Army for acquisition, logistics and technology told reporters this week. “That also aligns with the fact that the Army is probably the first service to look at [creating a program of record] for AI. So there’s been a lot of interest from the Office of the Secretary of Defense and other services that are asking, ‘Hey, are there things that we can work together on?’ and glean lessons learned.”

Commercial models for Army AI

Officials say wherever it makes sense, they want to adopt AI algorithms and models that industry has already developed. But there are nuances the Army needs to think through first.

For instance, the sorts of data the Army deals with — particularly in battlefield applications — is sometimes qualitatively different than the examples commercial models have been trained on. One example officials offered: An algorithm that analyzes imagery might have been trained on picture-perfect 4K video, but the Army needs it to also work with images taken from satellites, with perhaps some intermittent cloud cover.

“Another one of the obstacles for the adoption is how do we look at risk around AI? We have to look at issues around poisoned datasets, adversarial attacks, trojans and those types of things,” Bang said. “That’s easier to do if you’ve developed it in a controlled, trusted environment that DoD or the Army owns, and we’re going to do all that. But this really looks at how we can adopt third-party or commercial vendors’ algorithms right into our programs, so that we don’t have to compete with them. We want to adopt them.”

Bang said the Army is creating a “layered” risk reduction framework that it will apply to future commercial AI applications, with input from commercial vendors. Officials say they’ve met with more than 250 companies so far to discuss AI adoption.

Project Linchpin to build Army’s AI infrastructure

And for the most part, the service’s vehicle for buying those capabilities will be Project Linchpin, the new program of record the Army has been working to stand up over the past year.

Bharat Patel, the product lead for Linchpin at the Program Executive Office for Intelligence, Electronic Warfare and Sensors (PEO-IEWS), said the Army is treating the project as the AI “delivery mechanism” for its acquisition programs.

“It’s literally all the boring parts of AI: your infrastructure, your standards, your governance, your process. All those areas are things that we’re taking on, because that’s how you can tap into the AI ecosystem and deliver these capabilities at scale,” he said. “And as part of standing up the program, we’re doing a ton of market research. Since November of 2022, we’ve released four requests for information. We have well over 500 data points from all of those RFIs. They’re around fundamental things that we just have to do: How do I understand risk? How am I going to evaluate not only the algorithm, but how do I evaluate best of breed? How do I evaluate the company itself as part of the acquisition process?”

One of the most significant awards the Army has made so far under Project Linchpin was last October, when PEO-IEWS awarded up to $2 million to Booz Allen Hamilton and Red Hat to help develop “Traceability, Observability, Replaceability, and Consumability” (TORC) for the Army’s nascent AI infrastructure.

“Those are strategic design principles that we wanted to flesh out a little bit further to help us wrap security around the models, engage with industry, and create modular open system architecture concepts that will allow us to plug and play different services across the pipeline,” Patel said. “In order for you to get after AI, you have to have some kind of way to manage and label data, you have to have environments for model training … and then on top of it, all of that is different whether you’re talking text or [radio frequencies] or imagery. All of those are slightly different tools, slightly different processes, slightly different methods and approaches. Our TORC principles were a way for us to manage and evaluate and assess the various offerings so that we can plug and play different components and different companies.”

Using SBIR to reach small, nontraditional vendors

But the Army has already made numerous other smaller awards related to AI, largely through its Small Business Innovative Research (SBIR) and xTech programs. As of now, there are 115 active SBIR projects related to AI involving 75 different companies.

Dr. Matt Willis, the director of the Army SBIR Program, said many of those contracts were awarded in under 10 days.

“That really changes the value proposition for companies so they can actually get in and start contributing to our mission,” he said. “The Army needs to rely on the private sector to identify these best capabilities across the AI space, and in FY 2025, we’re predicting a significant investment in our SBIR program towards AI in particular, strategically aligned with Project linchpin. Those will be potentially up to or more than $150 million, which is about 40% of our SBIR program.”

Copyright © 2024 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.