A research-grade, reproducible benchmark and simulation platform comparing classical, optimal, game-theoretic, and learned missile-interception guidance.
Scope & ethics. Simulation-only, educational/research project using point-mass/kinematic models and public textbook algorithms. No hardware, no real targeting/sensor data, no munitions modeling, no detection-evasion tooling. See the README.
- Research foundation: Research Report — fact-checked state of the art across guidance, tracking, estimation, fusion, multi-agent, game theory, RL, optimization, and the identified gaps.
- Plan & roadmap: Project Plan — objectives → architecture → benchmark design → phases/milestones → risks → expected outcomes.
- Results digest: results.md — headline findings with the figure/experiment behind each.
- Benchmark methodology: benchmarks/methodology.md — how laws are compared, fairly and reproducibly.
- Algorithm notes: algorithms/ — governing equations and references for each guidance/estimation method.
Cross-paradigm guidance comparisons today are scattered across single studies on different engagement geometries and metrics, and most report only miss distance. INTERCEPT puts PN/APN, optimal/geometric, game-theoretic, MPC, and RL guidance on the same field — identical dynamics, shared scenario suite, shared metrics, full Monte-Carlo statistics — and makes the comparison reproducible.
Scenario (YAML) → Simulation Core (dynamics · RK4 · engagement loop)
├── Sensors → Estimation/Tracking (EKF/UKF/IMM) ┐
├── Guidance/Control (PN·APN·OGL·SMG·MPC·RL·Game) ├→ Engagement → Result/metrics
└── Adversary (scripted · game-theoretic · RL) ┘
Experiment & Analysis layer: Monte-Carlo · benchmark harness · capture-region · viz · RL training
The core is algorithm-agnostic; guidance laws, estimators, sensors, and adversaries are plug-ins conforming to a single controller/interface contract, so every paradigm runs against identical dynamics — the fairness property the benchmark depends on.
Simulation core. PointMass2D/3D dynamics, an RK4 integrator, line-of-sight geometry, the Entity/controller contract, and an Engagement loop with tunnelling-proof intercept detection. The same loop runs every fidelity level and both dimensions unchanged.
Guidance, six paradigms. Proportional Navigation (True/Pure/ZEM) and Augmented PN (notes); optimal LQ/ZEM (OGL), sliding-mode, and CasADi NMPC with an impact-angle constraint and event-triggered replanning (notes); Apollonius-circle geometric pursuit and a game-theoretic optimal evader (notes); and reinforcement learning (notes). Every law runs in 2-D and 3-D.
Fidelity ladder (L0→L3). AeroMissile2D/3D adds gravity, parasitic and induced drag, a g-limit, and autopilot lag; RealisticMissile2D/3D adds an ISA atmosphere, boost–sustain–coast propulsion with mass burn-off, Mach-dependent drag, and lift/dynamic-pressure-limited turning, so available g emerges from the physics rather than a prescribed limit. On realistic evasive engagements simple PN drops to 0.21–0.56 P(intercept) while prediction and robust laws recover to 0.79–1.00 (notes).
Estimation and sensing. Radar and IR-EO sensor models with seeded noise; EKF (Joseph), UKF, and IMM filters; and EstimatingGuidance closing the sense→estimate→guide loop in 2-D and 3-D (notes). The IMM holds ~9 m tracking error through a turn that diverges a single-model EKF to ~350 m.
The benchmark. A YAML scenario suite (2-D and 3-D), seeded Monte-Carlo built on the fairness invariant, Wilson-interval metrics, capture-region sweeps, and paired-bootstrap significance testing (methodology). A capstone heatmap spans paradigm × fidelity (L0→L3) × dimension with per-row significance.
Learned guidance that wins. From-scratch PPO collapses on the realistic plant; a residual parameterization (a bounded correction on a PN/APN baseline) restores it, and the recurrent APN-residual reaches P(intercept) 1.00/1.00/0.95 in 2-D — beating True PN (0.81) and Augmented PN (0.93) on the unpredictable jink — and full parity in 3-D (RL notes).
Many-vs-many and coordination. Hungarian weapon-target assignment with live re-assignment, a diverse-threat raid library, cooperative salvo (impact-time) and pincer-coverage guidance, and coordinated penetration tactics countered by an asset-value layered defense.
The INTERCEPT League. Every seeded engagement is a match; a Bradley-Terry fit puts every guidance law and every evader on one Elo ladder. Sliding-mode wins; the game-theoretic evader out-rates every guidance law.
Quality bar. 199 unit/property/regression tests passing, ruff-clean, fully type-hinted (with mypy running in CI). Guidance spans six paradigms; fidelity spans L0→L3; engagements run in 2-D, 3-D, and many-vs-many on one fair benchmark.
Open follow-ups: full PSRO self-play (Nash-mixture best response); decentralized MARL (MAPPO/PettingZoo); margin-aware League ratings (ordinal Bradley-Terry over miss distance).