haxballgym 1.0.0

Composable, batched RLGym-v2-style env over the headless haxball_core engine.

haxballgym

A composable, batched, RLGym-v2-style reinforcement-learning environment for Haxball, built on top of the headless haxball_core physics engine.

The physics is a faithful Rust port of Haxball's, verified to agree with the original engine to within 1e-9, and stepped in batches of thousands of matches in parallel — fast enough for serious self-play on a single machine.

Install

pip install haxballgym

This pulls in haxball_core (the Rust engine) automatically.

Quick start

from haxballgym import make_default_env

env = make_default_env(n_envs=512)
obs = env.reset()
obs, rewards, terminated, truncated = env.step(actions)

Or compose your own environment from the pieces — engine plus obs / action / reward / done / state-mutator builders:

from haxballgym import (
    TransitionEngine, DefaultObs, DiscreteAction,
    CombinedReward, GoalReward, TouchReward,
    GoalCondition, TimeoutCondition, KickoffMutator, Env,
)

See the project repository for the full docs, runnable training examples, and tools to play your trained policy back in the real browser game.

License

MIT