Multi-Agent Reinforcement Learning Environment for Beyond-Visual-Range Air Combat

An open simulation environment is presented for training multiple reinforcement-learning agents in beyond-visual-range air combat while balancing physical fidelity against rollout speed. The objective is to provide a reproducible testbed that retains the sensing, tracking, weapon, and cooperative-information effects governing long-range engagements without the computational cost of a full mission simulator. Aircraft and missiles use geodetic four-degree-of-freedom kinematics and energy-based aerodynamics; probabilistic radar detections, electronic-warfare effects, shared-track fusion, seeker-constrained missile guidance, and distance-dependent terminal lethality are represented explicitly. Fixed-shape observations and safety-gated actions support parallel self-play. End-to-end experiments with a memoryless policy learned boundary compliance, survival, and progressively more effective weapon employment, while subsystem profiling measured approximately ten seconds of wall-clock time for a 300-step two-versus-two episode on the reported hardware. These results establish internal training readiness and reproducible computational scaling and form a platform for reproducible research on learned tactics and fair benchmarking in this domain.