What would qualify as a step input on the plant so I can perform
If the system response is linear compared to the system input. For example: if I applied a step input with a input of 1 and got a 5 deg angle on the balance beam then applied a step input of 2 and got a 10 deg angle (and so on), then to capture system dynamics only one step input would be needed.
Conversely if I applied a step input with a input of 1 and got a 5 deg angle on the balance beam then applied a step input of 2 and got a 15 deg angle and a step input of 3 had a 25 deg angle then to capture the system dynamics all amplitudes should be tested.
Would it be a step command of X voltage applied to one motor? Would it need to be applied to both motors?
If each motor is controlled independently then yes, you would need to test one input at a time. It also matters if the SysID is a MIMO system, if the motor control is not independent, then you'll need to put in a step input, and then a step input in the negative direction.
It would also be advantageous to turn one motor on, then let the system return to normal, then turn the other on, let the system return to normal. Then turn one on then the other on at the same time. If the system goes unstable by turning one motor on, then you might not be able to get enough data for a system ID.
Do the units of plant input and output have to match?
No, this is taken care of by any gain parameters.
Make sure you get the order right, it would probably be advantageous to model the system on paper and see what kind of order it has and then use that order if using a linear model. If I remember right a simple inverted pendulum control system is non-linear so either use a nonlinear model or linearize the control inputs if this applies to the system