Table of Contents

Gym.NET documentation

Gym.NET gives .NET reinforcement-learning environments with Farama Gymnasium's API, and holds them to Gymnasium's behavior bit for bit and pixel for pixel. These pages cover the 1.0 API.

Guides

Page For
Getting started Making envs, the reset/step loop, seeding, spaces, render modes, wrappers, vector envs
Creating custom environments Writing an Env<TObs, TAct>, registering it for Gymnasium.Make, checking it with EnvChecker.CheckEnv
Migrating from 0.2 to 1.0 What changed in each 0.2 call, and Gym.Legacy for code that cannot move yet
Development and testing Building, the unit suite, the live parity suite against Gymnasium, goldens, [OpenBugs]

Reference

Page Holds
Architecture The packages and projects, the env contract, typed and untyped APIs, array ownership, how parity with Gymnasium is proven
Release notes 1.0.0 (1.0.0+rc1, release candidate 1) What 1.0 contains and leaves out, how to upgrade from 0.2, every known difference from Gymnasium, the ImageSharp advisories
Third-party notices Gymnasium (MIT), SDL_gfx (zlib), CPython (PSF), pygame-ce as a behavior reference
Example: CartPole Learning CartPole with the cross-entropy method, 100 carts at once in a vector env
Examples One project per environment the showreel plays (24), each playing its env with an agent: headless, in a window, or from the keyboard

Plans

The working documents behind 1.0, which they call 2.0, its planned number until 2026-11-12 (decision D23). They record why the code is the way it is; the pages above describe what it does.

Plan Holds
Completion plan The phases to 1.0, each with its status, scope and deferrals; the decisions D2–D23
Design review The approved implementation designs DR-01…DR-95, and the consequences found while implementing them
Gymnasium port plan The per-area specifications, the name map, the defect register
Parity testing The design of the live parity suite
NumSharp migration The move from NumSharp.Lite to NumSharp
MuJoCo assessment How MuJoCo (P9, resumed by DR-61) came to run on Gym.NET's own engine, measured: the 22 ids' reach into MuJoCo 3.15.0, why bitwise parity holds on win-x64, the steps and the decisions to review, a from-scratch HalfCheetah core that is bit-exact over 80,000 physics steps (§10), and line-coverage estimates for Humanoid, Hopper and the rest of the five most-used envs (§10.6)
Parity audit Each env's parity with Gymnasium by criterion, measured from both test suites (2026-11-11, regenerated by tools/parity/audit/parity_matrix.py), and the 2026-09 baseline before the port
C# code review Whether the design is sound C# (2026-11-11; its second edition measured again in Release): what is idiomatic, 23 findings ranked by value with their evidence (throughput against Gymnasium, profiles, fixes tried and verified), what removing each 0.2 leftover costs and what replaces it, and the NumSharp items upstream