Set the Rules, React to Change: Extending Natural-Language Drone-Swarm Coordination with Constraints, Events, and Operational Knowledge

Natural-language control can make drone swarms easier to direct, but one-shot commands are insufficient for missions with standing policies, changing priorities, and time-critical observations. We extend Swarm-Steward—a natural-language multi-robot coordination system that uses an LLM-based multi-agent architecture to ground operator requests and dispatch deterministic robot actions—with a mission-governance layer for human-supervised drone swarms. The system interprets operator intent and selects mission tools and predefined robot behaviors; drone motion, sensing, and interaction are executed through deterministic action primitives and controllers. Operators can define enduring constraints in ordinary language, such as “Never enter the helicopter landing zone” or “Keep all drones at least 10 m apart.” These instructions are translated into typed declarative rules and compiled into deterministic predicates over telemetry, map relations, planned trajectories, tool parameters, time, and action history. Constraints specify conditions that must remain satisfied and are validated against their semantics and the existing rule base before activation. They mediate LLM-selected plans before tool invocation, queueing, and execution, and during runtime monitoring, blocking violations or triggering replanning. Events use the same rule language and checking backend, but define operator-authored condition-response policies; when a condition becomes true, they can notify the operator, cancel or reprioritize queued actions, initiate replanning, or dispatch predefined robot actions. For example, “If any drone detects a person, stop its sweep and orbit the location” connects a perception event to a deterministic follow-up behavior without a new operator command. An Operational Knowledge Agent retains mission vocabulary, aliases, preferences, and reusable procedures, including conventions such as an operator’s preferred directional reference frame, enabling later instructions to be grounded consistently without modifying canonical mission entities.