[Question] How to realize Direct Env's _pre_physics_step function in Manager-based Env

[YuHoChau]

Hello, I’ve been working on training a locomotion policy for a quadruped that also has an arm (12 + 8 joints in total). Since this policy focuses solely on locomotion, the action space (12) is limited to the leg joint states. At each initialization, the arm is assigned a random fixed pose, which can be done easily using the Direct Env template as follows:

def _pre_physics_step(self, actions: torch.Tensor):
    # Store the raw actions
    self._previous_actions.copy_(self._actions)
    self._actions.copy_(actions)

    # Define leg joint indices in the robot's joint configuration
    leg_indices = [
        slice(0, 4), 
        slice(5, 9),
        slice(10, 14) 
    ]
    
    # Create a copy of the default joint positions
    self._processed_actions = self._robot.data.default_joint_pos.clone()
    
    # Only apply actions to the leg joints
    action_idx = 0
    for leg_idx in leg_indices:
        joint_count = leg_idx.stop - leg_idx.start
        self._processed_actions[:, leg_idx] = self._robot.data.default_joint_pos[:, leg_idx] + self.cfg.action_scale * self._actions[:, action_idx:action_idx+joint_count]
        action_idx += joint_count
        
def _apply_action(self):
    # Define arm joint indices based on joint names order
        # ['FL_hip_joint', 'FR_hip_joint', 'RL_hip_joint', 'RR_hip_joint', 'widow_waist', 'FL_thigh_joint', 
        # 'FR_thigh_joint', 'RL_thigh_joint', 'RR_thigh_joint', 'widow_shoulder', 'FL_calf_joint', 'FR_calf_joint', 
        # 'RL_calf_joint', 'RR_calf_joint', 'widow_elbow', 'widow_forearm_roll', 'widow_wrist_angle', 
        # 'widow_wrist_rotate', 'widow_left_finger', 'widow_right_finger'], 

    arm_indices = [4, 9, 14, 15, 16, 17, 18, 19]  # widow_waist(4), shoulder(9), elbow(14), forearm_roll(15), wrist_angle(16), wrist_rotate(17), fingers(18,19)
    # give a random action to arm joints
    self._processed_actions[:, arm_indices] = self._arm_random_initial_positions

    self._robot.set_joint_position_target(self._processed_actions)

However, how can I achieve similar functionality using the Manager-based environment? It appears that the ActionCfg (as shown below) cannot define or configure the joint states that are excluded from the action space:

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

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

More simple question would be, how can I convert my first part of the code from Direct Env to Manager-based Env? Thanks for your help!

[RandomOakForest]

Thank you for posting this. We have noted we haven't been able to reply to this as soon as we would prefer. The team will engage soon. Thank you for your interest in Isaac Lab.