I think Roger started with a few drones activated and weapons set to a Laser tag/MILES equivalent to test their learning curve. Unless you're well along the testing protocol, you start slow and work your way up. And when working with a system that "talks" to the individual units and learns from its mistakes and failures, you MUST have an end of exercise command or kill switch/command to terminate the exercise if something does wrong or levels up too quickly.
I think the exercises started simple and easy, but because of Roger 'winning' to often or "cheated" (Bang! Your dead! "No I'm not, you missed!) so the drones thought, "Well, if level one doesn't work let's up it to two...three...four an so forth." It possibly thought if one of two drones aren't enough, let's add one/two/three...all the drones until it got out of hand.
Roger is now hoping he can pull the plug on this exercise somehow before the drones catch him and "terminate with extreme prejudice" the target...him!