Implementation:Farama Foundation Gymnasium Transform Action Wrappers

Knowledge Sources	Farama_Foundation_Gymnasium Gymnasium Docs
Domains	Reinforcement_Learning, Wrappers
Last Updated	2026-02-15 03:00 GMT

Overview

A collection of action transformation wrappers that modify actions before they are passed to the underlying environment, including TransformAction, ClipAction, RescaleAction, and DiscretizeAction.

Description

This module provides four action wrappers built on a common TransformAction base:

• TransformAction -- Applies a user-provided function to the action before passing it to the environment's step function. Accepts an optional new action space specification.

• ClipAction -- Clips continuous actions to the environment's original action space bounds. The wrapper's own action space is set to infinite bounds (Box(-inf, inf)), allowing agents to output any value which gets clipped to valid bounds.

• RescaleAction -- Affinely rescales a Box action space to a new range [min_action, max_action]. Uses the rescale_box utility to compute gradient and intercept for the linear transformation.

• DiscretizeAction -- Uniformly discretizes a continuous Box action space into a Discrete or MultiDiscrete space. Computes bin centers for each dimension and maps discrete indices back to continuous actions. Supports both flattened Discrete (single integer) and MultiDiscrete (one index per dimension) modes.

Usage

Use these wrappers when you need to transform the action space: ClipAction for safety-clipping continuous actions, RescaleAction for normalizing action ranges to agent-friendly bounds, DiscretizeAction for using discrete algorithms on continuous environments, or TransformAction for custom transformations.

Code Reference

Source Location

• Repository: Farama_Foundation_Gymnasium
• File: gymnasium/wrappers/transform_action.py

Signature

class TransformAction(gym.ActionWrapper[ObsType, WrapperActType, ActType], gym.utils.RecordConstructorArgs):
    def __init__(self, env: gym.Env[ObsType, ActType], func: Callable[[WrapperActType], ActType], action_space: Space[WrapperActType] | None): ...

class ClipAction(TransformAction[ObsType, WrapperActType, ActType], gym.utils.RecordConstructorArgs):
    def __init__(self, env: gym.Env[ObsType, ActType]): ...

class RescaleAction(TransformAction[ObsType, WrapperActType, ActType], gym.utils.RecordConstructorArgs):
    def __init__(self, env: gym.Env[ObsType, ActType], min_action: np.floating | np.integer | np.ndarray, max_action: np.floating | np.integer | np.ndarray): ...

class DiscretizeAction(TransformAction[ObsType, WrapperActType, ActType], gym.utils.RecordConstructorArgs):
    def __init__(self, env: gym.Env[ObsType, ActType], bins: int | tuple[int, ...], multidiscrete: bool = False): ...

Import

from gymnasium.wrappers import TransformAction, ClipAction, RescaleAction, DiscretizeAction

I/O Contract

Inputs

Name	Type	Required	Description
env	Env	Yes	The environment to wrap (must have Box action space for Clip/Rescale/Discretize)
func	Callable	Yes (TransformAction)	Function to apply to the action
action_space	Space or None	No (TransformAction)	Updated action space for the wrapper
min_action	float, int, or ndarray	Yes (RescaleAction)	Minimum values for the rescaled action space
max_action	float, int, or ndarray	Yes (RescaleAction)	Maximum values for the rescaled action space
bins	int or tuple[int, ...]	Yes (DiscretizeAction)	Number of bins per dimension for discretization
multidiscrete	bool	No (DiscretizeAction)	If True, use MultiDiscrete instead of flattened Discrete (default False)

Outputs

Name	Type	Description
action	ActType	The transformed action passed to the underlying environment

Usage Examples

import numpy as np
import gymnasium as gym
from gymnasium.wrappers import ClipAction, RescaleAction, DiscretizeAction

# ClipAction: clip out-of-bounds actions to valid range
env = gym.make("Hopper-v4", disable_env_checker=True)
env = ClipAction(env)
_ = env.reset(seed=42)
_ = env.step(np.array([5.0, -2.0, 0.0], dtype=np.float32))
# The action [5.0, -2.0, 0.0] is clipped to [1.0, -1.0, 0.0]

# RescaleAction: rescale action space to [0, 1]
env = gym.make("Hopper-v4", disable_env_checker=True)
env = RescaleAction(env, min_action=-0.5, max_action=0.5)
env.action_space  # Box(-0.5, 0.5, (3,), float32)

# DiscretizeAction: convert continuous to discrete actions
env = gym.make("Pendulum-v1")
env = DiscretizeAction(env, bins=10)
env.action_space  # Discrete(10)
obs, _ = env.reset(seed=42)
obs, *_ = env.step(3)  # pass a discrete action index

Related Pages

• Environment:Farama_Foundation_Gymnasium_Python_3_10_Runtime