Gymnasium Robotics

TorchWM supports Gymnasium Robotics environments through shared factories and environment catalogs for online world models. The adapter imports and registers gymnasium_robotics before calling gymnasium.make, which exposes every environment registered by the installed Gymnasium Robotics package, including the legacy MuJoCo v2/v3 task ids that Gymnasium moved out of the core package.

Install

pip install "torchwm[robotics]"

You can also install the package directly:

pip install gymnasium-robotics

World-model availability

Gymnasium Robotics ids are listed dynamically from Gymnasium’s registry after gymnasium_robotics is installed. The catalog exposes those ids to the online world-model families (Dreamer, PlaNet/RSSM, IRIS, DIAMOND, Genie, and DiT). I-JEPA/JEPA is intentionally excluded because it trains on image datasets rather than online Gymnasium environments.

Dreamer configuration

from torchwm import DreamerConfig

cfg = DreamerConfig()
cfg.env_backend = "robotics"
cfg.env = "HalfCheetah-v2"
cfg.gym_render_mode = "rgb_array"

The Dreamer wrapper stack still applies action repeat, action normalization, and time limits after the Robotics environment is converted to TorchWM image observations.

Factory usage

from torchwm import make_robotics_env

env = make_robotics_env(
    "HalfCheetah-v2",
    seed=0,
    size=(64, 64),
    render_mode="rgb_array",
)

make_robotics_env() returns a GymImageEnv, so observations follow TorchWM’s image contract: {"image": uint8[3, H, W]}. Lower-level helpers also list all installed Robotics ids and retry Gymnasium creation when a moved MuJoCo v2/v3 id requires Robotics registration.

Example script

TorchWM includes a compact Dreamer example for Robotics environments:

python examples/run_dreamer_robotics.py --env HalfCheetah-v2 --total-steps 2000
python examples/run_dreamer_robotics.py --list-envs

MuJoCo compatibility

The MuJoCo and generic Gym image factories also recognize the moved-environment error emitted by Gymnasium. If a legacy id such as HalfCheetah-v2 fails because Gymnasium says the v2/v3 MuJoCo environments moved to Gymnasium Robotics, TorchWM imports/registers gymnasium_robotics and retries automatically.

Troubleshooting

• Moved v2/v3 MuJoCo environments: install torchwm[robotics] or gymnasium-robotics.

• Renderer errors: keep render_mode="rgb_array" for image-based Dreamer runs and verify that the underlying Gymnasium Robotics environment can render locally.

• Newer Gymnasium ids: modern ids such as HalfCheetah-v5 can continue to use env_backend="mujoco" or env_backend="gym" when they are available in your Gymnasium installation.