Teaching drones to dogfight
Teams from Georgia Tech Research Institute and the Naval Postgraduate School met in the barren hills of Camp Roberts — a training base for the California National Guard — last January to test their dogfighting skills. Or, rather, to test the dogfighting skills of their swarms of unmanned aerial vehicles.
Each team sent 10 propeller-driven Zephyr drones skyward with instructions to attack any enemies encountered. Although the drones were physically identical, they used different autonomy logic and collaboration and communications software that had been developed by the two teams.
“The ability to engage a swarm of threat UAVs with another autonomous swarm is an area of critical research for defense applications,” Don Davis, division chief of GTRI’s Robotics and Autonomous Systems Branch, told Georgia Tech News.
The dogfights weren’t exactly real-world. For starters, the drones didn’t directly keep track of friendly and enemy aircraft with cameras, radar and sensors. Instead, a computer on the ground monitored the locations of aircraft using their onboard GPS and fed the location information to the drones.
Davis told GCN that eventually sensors and vision processing will be added, “but the goal of this experiment was to evaluate tactics rather than sensors.”
Also, the drones couldn’t actually shoot at each other. Instead, they could “declare” a shot when they were in position to attack an enemy aircraft, and a computer model calculated the probability of a hit. Enemy aircraft left the field when “hit.”
OK, so which team won? “Well, we were not competing so much as we were cooperating — meaning that we decided together which tactics to run against each other, and we would deliberately pit a ‘simpler’ tactic against a stronger one just to measure the difference in performance,” Davis said. “Believe it or not, it really was a draw.”
He said that in the first test, the tactics were fairly simple, such as letting each aircraft pursue and attack the closest enemy vehicle. In subsequent tests, variations on those tactics included methods for more efficient target assignment so that resources were not wasted by having two or three vehicles pursue a single enemy, leaving others unprotected.
According to Davis, the drones that were given more complex tactics made better use of resources.
Among the most important outcomes, the test showed that an autonomous team can be used effectively in air-to-air engagements of an enemy swarm, he said. “We have developed powerful simulation tools to evaluate and evolve new tactics going forward,” he added. “Showing that flight test results align well to our simulations gives us confidence in the use of those tools.”
Both teams followed up with an exercise in April that involved students from the Air Force, Army and Navy. The Defense Advanced Research Projects Agency’s Service Academies Swarm Challenge fielded as many as 60 drones simultaneously, with a mix of fixed-wing and rotorcraft vehicles.
According to Davis, the challenge allowed the teams to test more complex tactics. “For example,” he said, “the rotorcraft could stay back in a defensive posture to guard a base while the fixed-wing vehicles could press forward to attack the enemy.”
Posted by Patrick Marshall on Aug 31, 2017 at 11:29 AM