README.md

metadata

license: mit
pretty_name: OpenFront.io game-state snapshots
tags:
  - reinforcement-learning
  - games

OpenFront.io game-state snapshots

250 bot/nation-only games of OpenFront.io (25 games on each of 10 maps), played headless on the real deterministic game engine (openfrontio/OpenFrontIO) and snapshotted every 10 ticks / 1s of game time (~375k snapshots total).

Generated by and documented at github.com/djmango/openfront-ai — see the repo for loaders, training code, and results.

Format v3: every snapshot captures the complete game state — the tile grid plus all entities (players, diplomacy, units, attacks in flight) — so a single snapshot is enough to reconstruct everything an agent could observe. The 10-tick cadence means a typical nuke flight (20-60 ticks) spans 2-6 snapshots, and units in transit carry their destination.

One tar per map. Each game directory contains:

• terrain.bin — uint8[w*h] immutable terrain (bit 7 land, bit 6 shoreline, bit 5 ocean, bits 0-4 magnitude; magnitude 31 = impassable)
• states/t<tick>.bin.gz — gzipped little-endian uint16[w*h] mutable tile state: owner smallID in bits 0-11 (0 = unowned), fallout bit 13, defense bonus bit 14
• states/t<tick>.json.gz — gzipped JSON with full entity state:
  • players: id (smallID), name, type, troops, gold, tiles, alive, traitor, disconnected, targets, embargoes, pending alliance requests (in/out), sparse relation list
  • alliances: [playerA, playerB, expiresAtTick] for each active alliance
  • units: type (city, port, warship, transport, silo, SAM, nukes in flight, ...), owner, x, y, target tile tx/ty (null if idle — for nukes this is the impact point), samLock (a SAM is intercepting), level, health, constructing, cooldown, troops
  • attacks: from, to (0 = terra nullius), troops, retreating, srcX/srcY (boat-landing origin, null for land attacks)
• meta.json — format version, map, dims, snapshot tick list, winner

A game evolving (each color a player, Onion map, ticks 3175 and 12375):

Generated by djmango/openfront-ai for training observation encoders for RL agents.