Public API

Simulusion exposes simulusion as the friendly public namespace for both application code and direct component imports. Use it for factory helpers, model classes, config classes, and environment constructors.

Common Workflow

python

import simulusion

# Discover supported factories.
models = simulusion.list_models()
backends = simulusion.list_env_backends()

# Configure and create a model in one step.
agent = simulusion.create_model(
    "dreamer",
    env_backend="dmc",
    env="walker-walk",
    total_steps=1_000_000,
)

# Create standalone environments through a consistent backend selector.
env = simulusion.make_env("CartPole-v1", backend="gym")

Factory Helpers

Helper	Purpose
`create_config(model, **overrides)`	Build the default config for a model family and apply validated overrides.
`create_model(model, config=None, **overrides)`	Instantiate a model or high-level agent by name. Config overrides are applied before construction; unknown fields are passed as constructor arguments when appropriate.
`make_env(env_id, backend="auto", **kwargs)`	Create an environment through a named backend such as `gym`, `atari`, `mujoco`, `robotics`, `brax`, or `unity`.
`list_models()`	Return canonical names accepted by `create_model`.
`list_env_backends()`	Return backend names accepted by `make_env`.
`list_envs(model=None)`	Return known environment IDs, optionally filtered by model family.

Direct Imports Still Work

The factory API is a convenience layer. Advanced and research workflows can keep using direct imports from simulusion:

python

from simulusion import DreamerAgent, DreamerConfig, RSSM

cfg = DreamerConfig()
cfg.env = "walker-walk"
agent = DreamerAgent(cfg)

Use direct imports when you are composing custom modules, subclassing internals, or need access to implementation-specific constructors.

1.0 scope

These names stay in the 1.x public surface, with the following documented limits:

There is no DreamerV3 implementation in 1.0, and no dreamer-v3 registry name: the Dreamer family is dreamer-v1 and dreamer-v2.
agent.train() with a step budget is the Dreamer-family path. Other registered models train through their dedicated trainers or CLI commands.
Genie VideoDataset loads .npy / .pt clips, or video files when the viz extra (OpenCV) is installed. TinyWorlds training still goes through scripts/train_genie_tinyworlds.py.

Edit this page on GitHub

Updated Oct 11, 2026