cambrian.utils.wrappers

Wrappers for the MjCambrianEnv. Used during training.

Classes

MjCambrianSingleAgentEnvWrapper	Wrapper around the MjCambrianEnv that acts as if there is a single agent.
MjCambrianPettingZooEnvWrapper	Wrapper around the MjCambrianEnv that acts as if there is a single agent, where
MjCambrianConstantActionWrapper	This wrapper will apply a constant action at specific indices of the action
MjCambrianTorchToNumpyWrapper	Wraps a torch-based environment to convert inputs and outputs to NumPy arrays.

Functions

make_wrapped_env(config, wrappers[, seed])

Utility function for creating a MjCambrianEnv.

Module Contents

class MjCambrianSingleAgentEnvWrapper(env, *, agent_name=None, combine_rewards=True, combine_terminated=True, combine_truncated=True)

Bases: gymnasium.Wrapper

Wrapper around the MjCambrianEnv that acts as if there is a single agent.

Will replace all multi-agent methods to just use the first agent.

Keyword Arguments:

agent_name – The name of the agent to use. If not provided, the first agent will be used.

action_space

Return the Env action_space unless overwritten then the wrapper action_space is used.

observation_space

Return the Env observation_space unless overwritten then the wrapper observation_space is used.

reset(*args, **kwargs)

Uses the reset() of the env that can be overwritten to change the returned data.

step(action)

Uses the step() of the env that can be overwritten to change the returned data.

class MjCambrianPettingZooEnvWrapper(env)

Bases: gymnasium.Wrapper

Wrapper around the MjCambrianEnv that acts as if there is a single agent, where in actuality, there’s multi-agents.

SB3 doesn’t support Dict action spaces, so this wrapper will flatten the action into a single space. The observation can be a dict; however, nested dicts are not allowed.

reset(*args, **kwargs)

Uses the reset() of the env that can be overwritten to change the returned data.

step(action)

Uses the step() of the env that can be overwritten to change the returned data.

property observation_space: gymnasium.spaces.Dict

SB3 doesn’t support nested Dict observation spaces, so we’ll flatten it. If each agent has a Dict observation space, we’ll flatten it into a single observation where the key in the dict is the agent name and the original space name.

property action_space: gymnasium.spaces.Box

The only gym.Space that SB3 supports that’s continuous for the action space is a Box. We can assume each agent’s action space is a Box, so we’ll flatten each action space into one Box for the environment.

Assumptions:

• All agents have the same number of actions

• All actions have the same shape

• All actions are continuous

• All actions are normalized between -1 and 1

class MjCambrianConstantActionWrapper(env, constant_actions)

Bases: gymnasium.Wrapper

This wrapper will apply a constant action at specific indices of the action space.

Parameters:

constant_actions – A dictionary where the keys are the indices of the action space and the values are the constant actions to apply.

step(action)

Uses the step() of the env that can be overwritten to change the returned data.

class MjCambrianTorchToNumpyWrapper(env, *, convert_action=False)

Bases: gymnasium.Wrapper

Wraps a torch-based environment to convert inputs and outputs to NumPy arrays.

step(actions)

Using a numpy-based action that is converted to torch to be used by the environment.

Parameters:

action – A numpy-based action

Returns:

The numpy-based observation, reward, termination, truncation, and extra info

reset(*, seed=None, options=None)

Resets the environment returning numpy-based observations and info.

Parameters:

• seed – The seed for resetting the environment

• options – The options for resetting the environment

Returns:

The numpy-based observation and extra info

render()

Renders the environment returning a numpy-based image.

Returns:

The numpy-based image

make_wrapped_env(config, wrappers, seed=None, **kwargs)

Utility function for creating a MjCambrianEnv.