Despite the near-ubiquitous discussion in the Department of Defense surrounding autonomy, many in the operational and acquisition community acknowledge that much work remains about how machines will make decisions.

“The autonomous systems that we have today are what we would call directive autonomy, where we go into mission planning and we tell the system exactly where it should go or what parameters it should operate within and we don’t allow it to make decisions outside those boundaries,” Rear Adm. Mark Darrah, Program Executive Officer for Unmanned Aviation and Strike Weapons, said during an April 11 panel at the Navy League Sea Air Space conference.

“When we start talking about moving beyond the current directive autonomy we put into our systems, will we start seeing stochastic type behaviors you would get with machine learning, how are we going to program that in with mission planning,” he said. “We tend to focus very heavily on the machine itself and its decision making and so on but we have to set the parameters for what it will base its decisions on. A lot of work that needs to be done there.”

Darrah also discussed where to insert autonomy in the kill chain.

“Right now we have a lot of legacy systems that have been optimized to operate a certain way. We put autonomy in there, it could be like having a Lamborghini and only being allowed to drive 20 miles an hour,” he said. “It can make very, very fast decision but if it’s reliant and interfaced with a legacy system that can’t make the movements on those quick decisions then we haven’t benefited our kill chains.”