system identification

System identification means measuring the real system to figure out its true parameters — masses, friction, motor constants, delays — so that your simulator matches reality more closely. It attacks the reality gap from the simulator side, rather than from the policy side.

You run experiments on the real hardware, record states and actions, and then fit the simulator parameters that best reproduce the observed behaviour, usually by minimizing prediction error. In RL this can be done once up front, or online — continually re-estimating the dynamics as the robot operates and adapting the policy to the refined model. It complements domain randomization nicely: identify what you can, and randomize what you cannot.