A fully observable multi-domain Capture The Flag combat simulation environment for reinforcement learning.

Author: Sanjna Ravichandar

    sanjna@codemetal.ai

Repository Link: ‣

For any bugs or suggestions, please open an issue in Github.

Contents:

1. Game Overview
2. Agent Objective
3. Action Space
4. Observation Space
5. Reward
6. Training Overview
7. Key points for Development

📘 Overview

FlagFrenzyEnv is a custom Gymnasium-compatible environment simulating a dynamic military scenario between two factions—Legacy (Blue Team) and Dynasty (Red Team)—in a large, continuous 2D battlefield (1,300,000 × 1,300,000 km). The agent assumes the role of the Blue Team and learns to command entities in real-time to achieve mission objectives through tactical movement, engagement, and retreat decisions.

Game Premise

• Team Blue (Legacy): Controlled by the agent.
  • Controllable packages: MADDOG, B1
  • Uncontrollable package: AWACS
• Team Red (Dynasty): Controlled by scripted logic.
  • 6 controllable packages
  • 5 uncontrollable packages

Each package consists of 2–5 squadrons, and each squadron consists of 1–4 units. The game operates at the squadron level, and units are subject to domain-specific capabilities (e.g., air-to-ground strikes).