[Question] Question about applying actions and observations in the manager base navigation.

[celestialdr4g0n]

Why there is no set_joint_velocity_target() or set_joint_position_terget() in PreTrainedPolicyAction class in the navigation example? Is actions in the navigation example fully comes from the pre-trained policy?

class PreTrainedPolicyAction(ActionTerm):
    """
    Operations.
    """

    def process_actions(self, actions: torch.Tensor):
        self._raw_actions[:] = actions

    def apply_actions(self):
        if self._counter % self.cfg.low_level_decimation == 0:
            low_level_obs = self._low_level_obs_manager.compute_group("ll_policy")
            self.low_level_actions[:] = self.policy(low_level_obs)
            self._low_level_action_term.process_actions(self.low_level_actions)
            self._counter = 0
        self._low_level_action_term.apply_actions()
        self._counter += 1

What is the output actions of the pre-trained file?
If the actions from the pre-trained checkpoint (joint positions maybe?) are assigned to self._raw_actions of the PreTrainedPolicyAction class, why did you remap the self._raw_actions to "velocities_command" while the action of manager_based -> locomotion -> velocity_env is joint positions?

[Mayankm96]

The object _low_level_action_term handles the conversion to joint actions.

[celestialdr4g0n]

Thank you for your response.
Do you mean that the _low_level_action_term object will "call" the action term in manager_based -> locomotion -> velocity_env?