A model must survive another test
Estimating a plant from one response is the beginning of a model, not the end of validation.
First, picture it
A motor model might match a small speed step perfectly because you tuned its parameters to that trace. Will it also predict a different target or load pulse? That unseen trial is more revealing.
What the model says
System identification collects known inputs and measured outputs, chooses a model structure, estimates parameters and checks predictions on independent data. A step can suggest gain and dominant time scale, but the motor has electrical and mechanical states; a single noisy step may not identify all constants uniquely. In this lab the nominal motor is the design model, while inertia and friction multipliers alter the simulated actual plant. This is a controlled model-mismatch exercise, not an estimator fitted to uploaded hardware data. Compare with a held-out target or load condition before claiming the model is useful for control.
prediction error = measured output − model outputUse a separate validation run rather than scoring only the trace used to choose parameters.
Make it concrete
Pin the nominal motor at a fixed controller and target. Double actual inertia; the nominal Bode and root locus do not move, but the time response does. Change target or load to probe whether the mismatch matters for your objective.
Test a changed motorA good fit on the experiment used to select parameters is not independent evidence of predictive accuracy; neither is a pretty nominal stability margin.
Check your understanding+
Why test a model on a load pulse that was not used to tune it?
It checks predictive behavior outside the fitting trial and exposes errors relevant to disturbance rejection.