I don't have time to read the paper right now, but yeah generally feed-forward just means providing the signal you expect will give you the output you want. That would mean you have some advance knowledge of how to jockey the system to get the output you want, which is some kind of model. The model can be anything - differential equations, a lookup table, etc. Whatever you want that lets you go, "Okay, I need X input to get Y output."
The beautiful thing about feed-forward control is that you don't need an error term like you would for PID control. Because a PID control only acts on reference error the PID controller can only ever be reactive. Another issue with PID control is that it can be very difficult to adequately tune it to large differences in reference magnitudes or to large variations in reference shapes (ramp vs. step, etc.).
When you do feed-forward control you minimize the control error because your system is proactive - you start taking action now to meet the reference instead of waiting for the reference error to build up large enough to become actionable. Additionally, you minimize the magnitude of the reference error across a variety of scenarios because the feed-forward control does the "bulk" action and the feedback control exists to make up for the modeling errors or disturbance inputs. Getting to a more repeatable reference error means you can better tune your PID controller and that the gains you choose will be appropriate for a wider range of scenarios.
