Crate gymnasia 

A Rust implementation of OpenAI’s Gymnasium environments.

Architecture

Unlike Python Gymnasium, gymnasia separates simulation from rendering.

The crate is organized into layers:

1. core::Env — pure simulation. step() and reset() perform physics only. No feature flags, no rendering code, no graphics imports.

2. core::Renderable — a trait that produces a render::draw::DrawList (a backend-agnostic list of draw commands) from the current state. Always compiled — DrawList has zero graphics dependencies.

3. RenderEnv (requires render feature) — a wrapper that composes an Env + Renderable with a macroquad-backed screen. Implements Env so it participates in wrapper chains.

4. wrappers — composable behavior wrappers (TimeLimit, RecordEpisodeStatistics, NormalizeObservation, etc.) that implement Env via generic delegation.

This means the simulation compiles and runs with zero dependencies on any graphics library. Rendering is opt-in via --features render.

Quick start

use gymnasia::core::Env;
use gymnasia::envs::classical_control::cartpole::CartPoleEnv;

let mut env = CartPoleEnv::new();
env.reset(None, Default::default());
let result = env.step(1);

Modules

core

Core traits and types: Env, StepResult, Flatten, Renderable.

envs

Concrete environment implementations.

render

Drawing, rendering backend, and the RenderEnv wrapper.

spaces

Space descriptors: Discrete, BoxSpace, etc.

utils

Shared utilities: seeding, clip, etc.

wrappers

Composable environment wrappers. Composable environment wrappers.

Macros

delegate_env

Implements Env for a wrapper by delegating all methods to self.env.