Motrix Environment Core

motrix-env-core provides the backend-neutral framework APIs used to build reinforcement learning environments:

• environment base classes and configuration protocols;
• environment registry and registration decorators;
• backend-independent scene configuration and the simulator boundary (motrix_env_core.sim);
• NumPy environment runtime and rendering;
• framework-level math and reward utilities.

It intentionally contains no built-in tasks, robot models, meshes, textures, or motion data, and no simulator: importing motrix_env_core never imports MotrixSim or MuJoCo. Install motrix-envs when the built-in MotrixLab environments are required.

from motrix_env_core import EnvCfg, SceneCfg, configclass, registry

Environment implementation packages register themselves by importing their configuration and environment modules. The framework does not discover or import any concrete environment package.

Simulator backends

Simulator bindings live in their own distributions and register through the motrix_env.sim_backends entry-point group:

• motrix-env-motrixsim — the MotrixSim backend (live simulation, rendering, and the torch-tensor frontend); registered as the default simulator;
• motrix-env-mujoco — a compile-only MuJoCo backend that lowers SceneCfg into mujoco.MjModel (sim2sim, preview, conversion tools).

Installing a backend package registers it; the registry discovers entry points lazily on first use and never imports a native simulator until a backend is actually constructed.

Compile-only consumers (viewers, converters, previews) import the backend's scene compiler directly, for example from motrix_env_mujoco.compiler import MuJoCoSceneCompiler.

Requesting a backend whose package is not installed fails loudly with an installation hint, and the other registered backends remain usable.