�Ӱ���Ƶ

When it comes to developing technology in the fast-paced world of artificial intelligence (AI), speed is critical to delivering national defense solutions. That’s why a team at the Johns Hopkins Applied Physics Laboratory (APL) in Laurel, Maryland, is designing a virtual simulation that can expedite the advancement of more robust and effective AI for the warfighter.

Through Operation Protovision, a four-event competition sponsored by the Air Force Research Laboratory (AFRL) as part of its , APL is leading eight organizations on an accelerated journey to design a new generation of Networked, Collaborative and Autonomous (NCA) weapons systems. The first of the four competitions is set for Dec. 8 and 9.

Core to the competition, and a key enabler of its speed and scale, is an APL-built simulation environment, the first prototype of the fully digital portion of what will be a new live, virtual, constructive capability for weapons research and development. Dubbed “the Colosseum,” this virtual space allows the diverse range of competitors to test, evaluate and develop algorithms for an NCA system — in this instance, cruise missiles — in a digital approximation of the real world.

“The main design challenge is constructing the Colosseum to seamlessly marry the various competitors and their tech under evaluation against the challenge problem,” said Brandon Coloe, a guidance, navigation and control systems engineer in APL’s Force Projection Sector (FPS) and the project’s technical lead. “We want to lower the barrier to integration as much as possible, so we can pull in as many different competitors as we can to really get the best tech out of the other end.”

APL is serving as facilitator and innovator, collaborating with AFRL as it uses the Pentagon’s Defense Innovation Unit to rapidly bring nontraditional technology developers with great ideas into the defense fold. To ensure each organization — regardless of size or background — can plug into a Colosseum that presents a meaningful challenge problem, the team leveraged APL’s large portfolio of work in advanced autonomy for this latest problem set.

“We previously applied our artificial intelligence, machine learning and reinforcement learning experience to the air domain, when we created a similar competition environment for the 2020 Defense Advanced Research Projects Agency AlphaDogfight trials,” explained APL’s Christopher DeMay, who manages the FPS Intelligent Combat Platforms program. “Now APL is expanding the capability to include the weapons domain.”

Designing the "Challenge Problem"

Shifting from fighter jets to collaborative weapons, however, brings a new set of complexities and a multitude of factors.

“The challenge problem is a stressing scenario against a regional air defense system,” explained APL’s mathematician- and physicist-trained Robert Shearer. “It contains both long- and short-range missile defense systems, as well as early warning radars and jamming systems, all split into sectors under control by a sector commander.”

The task for Shearer, alongside APL engineer Steven VanDerwalker, was to develop this representative air-to-ground strike scenario in AFRL’s Advanced Framework for Simulation, Integration and Modeling (AFSIM), which would serve as the basis for the competition. The duo not only turned the scenario into a working modeling and simulation construct but also fine-tuned that construct to be as meaningful as possible.

“Early in the process we had conversations with AFRL and the Defense Innovation Unit about aspects of the air-to-ground strike missions to stress in our scenario,” explained VanDerwalker. “Dynamic target assignment and dynamic path planning were identified as ideal candidates for machine-learning techniques to enhance cruise missile effectiveness, so from there we made sure to place opportunities for the machine learning to stress those aspects in our scenario.”

Building the Colosseum

The team then had to develop the Colosseum software architecture to allow the various competitors and their constructed autonomous algorithms to easily interact with the simulated challenge problem itself.

“Our role in developing a software interface is to lower the barrier to entry for the performers and vector their efforts toward developing meaningful, innovative solutions on the core problem as much as possible,” explained APL controls and optimization expert Nick Watkins.

The team built containerized architecture that allows each competitor’s algorithms to run separately and, in APL computer scientist Kyle Casterline’s words, “talk” to a common interface. That interface then controls the simulation step rate and shuttles data from agents to simulation and back again.

“You want that loop to be as efficient as possible,” said Casterline. “Every bit of time you save per simulation step adds up when you’re hosting hundreds to thousands of competition rounds.”

In addition to optimizing Colosseum efficiency, the team considered its scalability.

“A single simulation installs and runs fine on a computer, but running 1,000 instances of a simulation on a reasonable time scale introduces a number of technical problems,” explained Casterline. “We’re using virtualization tools to bundle the simulations and deploy them across a broader set of hardware. Once you have that software architecture in place, it’s a matter of just acquiring additional hardware to scale further and further.”

“The solutions always came down to writing well-thought-out software,” added Watkins.

The opening event, described by DeMay as the “preseason” competition, is the first real test of the interaction between the Colosseum challenge problem, interface and competitors.

“This is our first opportunity to field the team,” DeMay added. “The threshold is to deliver all agents and get them working in the simulation experience. As the esprit de corps and excitement builds, the objective will be to get after harder aspects of the technology, moving beyond functionality and toward performance. We want to challenge them to a point that lays the groundwork for the subsequent competitions.”

The next three competitions, all planned for 2022, will continue to challenge the competitors and their algorithms by incrementally increasing the fidelity and complexity of the scenario. As they do, each competition will move closer to the ultimate goal: beginning to transition winning technology from simulated environments to the real world — all in under a year.