evorl.envs.gymnasium

Module Contents

Classes

GymnasiumAdapter	Adapter for Gymnasium to support Gymnasium environments.
OneEpisodeWrapper	Vectorized one-episode wrapper for evaluation.

Functions

create_gymnasium_env	Create a gym env based on Gymnasium.
gymnasium_space_to_evorl_space

API

class evorl.envs.gymnasium.GymnasiumAdapter(env_name: str, max_episode_steps: int, num_envs: int, record_ori_obs: bool = False, discount: float | None = None, vecenv_kwargs: dict | None = None, **env_kwargs)

Bases: evorl.envs.env.EnvAdapter

Adapter for Gymnasium to support Gymnasium environments.

This env already is a vectorized environment and has experimental supports. It is not recommended to direcly replace other jax-based envs with this env in EvoRL’s existing workflows. Users should carefully check the compatibility and modify the corresponding code to avoid the side-effects and other undefined behaviors.

Caution

This env breaks the rule of pure functions. Its env state is maintained inside the gymnasium. Thesefore, users should use it with caution. Unlike other jax-based envs, this env has following limitations:

• No support for recovering from previous env state.

  • In other word, you can’t rewind after calling env.step.

  • For example, you can’t resume the training from a checkpoint exactly as before; Similarly, evorl.rollout.eval_rollout_episode will also result in undefined behavior.

• We use gymnasium’s AsyncVectorEnv, which uses python’s multiprocessing package for parallelism. This may cause performance issues, especially when the number of parallel environments is large.

  • We recommend that the total number of parallel environments does not exceed the number of CPU logic cores.

property action_space: evorl.envs.space.Space

property obs_space: evorl.envs.space.Space

reset(key: chex.PRNGKey) → evorl.envs.env.EnvState

setup_env_callback()

step(state: evorl.envs.env.EnvState, action: evorl.types.Action) → evorl.envs.env.EnvState

class evorl.envs.gymnasium.OneEpisodeWrapper(env: evorl.envs.env.Env)

Bases: evorl.envs.wrappers.Wrapper

Vectorized one-episode wrapper for evaluation.

step(state: evorl.envs.env.EnvState, action: evorl.types.Action) → evorl.envs.env.EnvState

evorl.envs.gymnasium.create_gymnasium_env(env_name, episode_length: int = 1000, parallel: int = 1, autoreset_mode: evorl.envs.wrappers.AutoresetMode = AutoresetMode.ENVPOOL, discount: float | None = 1.0, record_ori_obs: bool = False, **kwargs) → evorl.envs.gymnasium.GymnasiumAdapter

Create a gym env based on Gymnasium.

Unlike other jax-based env, most wrappers are handled inside the gymnasium.

evorl.envs.gymnasium.gymnasium_space_to_evorl_space(space: gymnasium.Space) → evorl.envs.space.Space