The model is only right in a perfect world
A dynamics model gives you the ideal torque needed to achieve a motion, but that calculation assumes the model's inputs are exact, the true mass, the true friction, the true center of gravity. Real robots never quite match this. Friction in a gearbox might be higher than estimated, a cable might catch, a payload might be slightly heavier than assumed, or the robot might simply get bumped mid-motion. None of these show up in the feedforward torque computed from the dynamics model alone.
Control theory is the discipline that deals with this gap. Instead of trusting the dynamics model's prediction blindly, a controller continuously measures the robot's actual state, position, velocity, sometimes force, and compares it to where the robot was supposed to be. The difference between intended and actual state is the error, and the controller's job is to compute a correction that drives that error back toward zero.