Topology-Safe Air-to-Air Interception via “Go”-Inspired Metrics and Multi-Agent Reinforcement Learning

Airspace defense against agile UAVs requires multiple defenders to cooperatively encircle evasive intruders, but existing formation controllers cannot certify when an intruder is truly trapped. This paper introduces a “Go”-inspired framework that maps the board-game concept of liberties to continuous-space escape corridors, yielding a Benson-inspired dual capture criterion: total liberty below a threshold AND sufficient directions blocked. A directional relevance gating mechanism enforces that interceptors behind the evader cannot contribute to blocking, naturally requiring multi-agent cooperation. The metrics feed a MAPPO-based learning layer with curriculum-scheduled capture thresholds under centralized training/decentralized execution. In 3-vs.-1 simulations with five independent seeds, trained interceptors achieve a 52.9±3.7% capture rate under deterministic evaluation against a heuristic evader under the strict multi-directional criterion, with evader liberty reduced to 0.36 and 5.1/8 directions blocked.