To prevent a slippery slope of prev_prev_prev_prev_..._action, I suggest we instead make a tensor for action history (N, H, N_a), where the action manager receives something like a history length in its configuration (similar to what we do for a contact sensor).

Where should I add history_length in ActionsCfg?

@configclass
class ActionsCfg:
    """Action specifications for the MDP."""

    joint_pos = mdp.JointPositionActionCfg(asset_name="robot", joint_names=[".*"], scale=0.5, use_default_offset=True)

Where should I add history_length in ActionsCfg?

@configclass
class ActionsCfg:
    """Action specifications for the MDP."""

    joint_pos = mdp.JointPositionActionCfg(asset_name="robot", joint_names=[".*"], scale=0.5, use_default_offset=True)

I think we can add a history_length attribute to the ActionTermCfg - https://github.com/isaac-sim/IsaacLab/blob/main/source/extensions/omni.isaac.lab/omni/isaac/lab/managers/manager_term_cfg.py#L77. then in the example above, the predefined action terms such as JointPositionActionCfg can all leverage it.

In ActionTerm, we don't have ActionGroup like ObservationGroup in ObservationTerm, so if I set history_length for every action terms separately, how should I store history for different action terms?

(If I add a history_length attribute to the ActionTermCfg, I can't get self.cfg.history_length such as self._action_history = torch.zeros(self.num_envs, self.cfg.history_length, self.total_action_dim, device=self.device). )

Ah I see what you mean, it is tricky indeed. I guess one way is to just use the max history length of all of the terms and populate the buffer for them all in the Action Manager, similar to how the current prev action is done. Or maybe we can use a special named term in the action manager cfg for specifying the history length? both seems a bit messy though.

@Mayankm96 what would you recommend? maybe I'm missing something in my understanding of the action manager.