Skip to content
SimulusionDocs

Search guides, algorithms and the platform.

Environment Backends

Environment Backends

Simulusion ships environment adapters for pixel-based world-model training, model-based reinforcement learning, and benchmark collection. Each backend page explains the installation requirements, factory functions, observation and action conventions, configuration fields, and common troubleshooting steps for that backend.

Choosing a backend

Backend Best for Primary APIs Typical observations Typical actions
DeepMind Control Suite Dreamer-style continuous-control tasks with state and rendered image observations DeepMindControlEnv, env_backend="dmc" Dict with DMC state keys plus image Continuous Box from the DMC action spec
DeepMind Lab 3D navigation and puzzle tasks from DeepMind Lab DMLabEnv, make_dmlab_env, env_backend="dmlab" Dict with image plus requested Lab observations Normalized one-hot Box[-1, 1] over native Lab actions
Gym and Gymnasium Classic control, Box2D, custom Gym environments, and generic rendered tasks GymImageEnv, make_gym_env, env_backend="gym" Dict with image only Original continuous Box or one-hot vector for discrete actions
DeepMind BSuite Small diagnostic RL benchmark tasks such as catch/0 and deep_sea/0 BSuiteImageEnv, make_bsuite_env, env_backend="bsuite" Dict with synthetic image, the flat state vector, or both One-hot vector for discrete actions
Minecraft Diamond DreamerV3's Minecraft benchmark: obtain a diamond from sparse milestone rewards MinecraftDiamondEnv, env_backend="minecraft" Dict with image plus inventory and status vectors One-hot vector over 25 actions
Brax JAX/Brax continuous-control tasks through a Gym-like image adapter BraxImageEnv, make_brax_env, env_backend="brax" Dict with synthesized image; raw vector in info["vector_observation"] Continuous Box[-1, 1] matching env.action_size
Atari Atari 2600 environments through Gymnasium/ALE make_atari_env, make_atari_vector_env ALE RGB/RAM observations Discrete Atari actions
Procgen Procedurally generated benchmark games ProcgenImageEnv, make_procgen_env, env_backend="procgen" Dict with image One-hot vector for discrete Procgen actions
MuJoCo Gymnasium MuJoCo task ids and native MJCF/MJB models make_mujoco_env Image dict via GymImageEnv/MuJoCoImageEnv Continuous Box
Gymnasium Robotics All ids registered by the installed Gymnasium Robotics package, including moved legacy MuJoCo v2/v3 ids make_robotics_env, list_gymnasium_robotics_envs Image dict via GymImageEnv Continuous Box
Unity ML-Agents External Unity executable simulations with continuous-control behaviors UnityMLAgentsEnv, env_backend="unity_mlagents" Dict with image Continuous Box[-1, 1]
Vectorized environments Multiprocess/vector rollout collection and native ALE vectorization TorchVectorizedEnv, make_atari_vector_env Batched observations Batched actions
World Model Env Model-based RL, policy optimization, and evaluation inside learned dynamics WorldModelEnv, make_world_model_env, env_backend="world-model" Adapter-defined Gymnasium space Adapter-defined Gymnasium space
Wrappers Shared preprocessing, action conversion, time limits, reward observations, and image transforms simulusion.envs.wrappers Backend-dependent Backend-dependent

Shared conventions

Most Simulusion training code expects image observations as a dictionary entry named image with channel-first shape (3, H, W) and uint8 values. Backend adapters that wrap vector-only environments synthesize an image representation so pixel-based agents can still run.

DIAMOND-style Atari support is documented on the Atari page as a preprocessing helper for Atari rollouts. It is not a separate environment backend.

Dreamer environment construction applies a standard wrapper stack after creating DMC, DMLab, Gym/Gymnasium, MuJoCo, Gymnasium Robotics, Procgen, BSuite, Minecraft, Brax, or Unity environments:

  1. ActionRepeat repeats each selected action for cfg.action_repeat environment steps.
  2. NormalizeActions exposes finite continuous action bounds as normalized [-1, 1] policy outputs.
  3. TimeLimit truncates episodes after cfg.time_limit // cfg.action_repeat wrapper steps.

Use simulusion envs list to print the lightweight backend catalog used by the CLI.

Edit this page on GitHub

Updated Nov 3, 2026