Drone-Megaddon 2026

Co-Intelligent Swarm Control with Language and Vision Models

Abstract

In 2014, we introduced Drone-Megaddon, an Android interface for managing drone swarms, patterned after real time strategy (RTS) video games like StarCraft II. We argued that its select-then-execute paradigm fit commanding multiple drones better than the single drone piloting interfaces available at the time. Twelve years later, three things around it have changed. Commercial drones now ship with LTE connectivity and with navigation modes built for environments where GPS is denied. Drone warfare at scale has generated a public body of operational lessons on electronic warfare and operator load. And foundation models have made natural language intent and visual grounding tractable as interfaces to robotic systems. The third change is the one this paper turns on. We propose a successor to Drone-Megaddon that keeps the 2014 core: RTS selection, the $RAD command queue, ACK’d radio, and human oversight. Over that core we propose three AI layers. The intent translator takes an operator utterance together with the currently selected drones and the map state, and emits a validated sequence of $RAD commands. The perception grounder takes a referring expression such as “follow the red sedan” together with the drone’s current video feed, and emits a tracked bounding box bound to a specific drone. The coordination planner takes a single waypoint intended for multiple drones and emits one distinct final position per drone, addressing what the 2014 paper called the random offset at destination behavior. We describe each layer at the level of its interface, walk through a search-and-rescue example, and address ethical considerations the 2014 paper did not engage with. This is a proposal; the detailed architecture is in §5.