Language

unilab.envs.mdp.actions.actions.BaseAction¶

class unilab.envs.mdp.actions.actions.BaseAction[source]¶

Bases: ActionTerm

Apply a cold-path-resolved affine transform to raw policy actions.

Parameters:

cfg (BaseActionCfg)
env (ManagerBasedRlEnv)

Methods

`__init__`(cfg, env)
`apply_actions`()
`process_actions`(actions)
`reset`([env_ids])	Resets the manager term.

Attributes

`action_dim`
`name`
`num_envs`
`offset`
`processed_action`
`raw_action`
`requires_substep_state_feedback`	Whether `apply_actions` needs state produced by the previous physics substep.
`scale`
`target_ids`
`target_names`
`cfg`

cfg: BaseActionCfg¶

__init__(cfg, env)[source]¶

Parameters:

cfg (BaseActionCfg)
env (ManagerBasedRlEnv)

property scale: float | ndarray¶

property offset: float | ndarray¶

property raw_action: ndarray¶

property processed_action: ndarray¶

property action_dim: int¶

property target_ids: ndarray¶

property target_names: list[str]¶

process_actions(actions)[source]¶

Parameters:: actions (ndarray)
Return type:: None

reset(env_ids=None)[source]¶

Resets the manager term.

Parameters:: env_ids (ndarray | slice | None)
Return type:: None

__call__(*args, **kwargs)¶

Returns the value of the term required by the manager.

Return type:: Any

abstract apply_actions()¶

Return type:: None

property name: str¶

property num_envs: int¶

requires_substep_state_feedback: ClassVar[bool] = False¶

Whether apply_actions needs state produced by the previous physics substep.

False declares that the generated control is invariant across all physics substeps in one control step, so the env may apply it once before a batched backend step. State-feedback terms must override this declaration with True.