What is the best practice to load a trained low-level policy using DirectRL workflow for a robot as controller?

[lei3li]

I trained a loco-manipulation policy for a robot using DirectRL workflow and want to use it as a low-level controller policy, where the high-level policy involves visual inputs and other types of information. What is the best practice to load the low-level policy as a controller for the robot to take in commands out of the high-level policy? I know there is a PreTrainedPolicyAction for ManagerBased workflow but I'm getting this information that it's not supported to train a policy with the types of Observation in my project, I'm also checking on the PolicyController. I want to know what is the best practice from your professional perspective to meet my demand for my tasks? Any advice would be appreciated!