Can the training be decoupled from RLbench/PyRep?

[IrvingF7]

trafficstars

RLbench/PyRep is not being actively updated to keep up with newer versions of Ubuntu, so it's basically stuck on Ubuntu 20, which is a 5-year-old EOF OS

Currently, the repo seems to deeply couple RLbench/PyRep with the training code to the point that you cannot run train.py without them. Currently, I am not able to deploy the code in an HPC setting because the headless mode of RLbench/PyRep is cumbersome (like requiring sudo access and rebooting the system, etc)

Can the author provide some pointers as to how to decouple the training/model from RLbench/PyRep? I assume this is needed for a real-world experiment anyway, right?