We show system identification of self-organizing, swarm robotic systems using a ``gray-box'' approach, based on probabilistic macroscopic models. At hand of a well studied case study concerned with the autonomous inspection of a regular structure by a swarm of miniature robots, we show how to achieve predictive models of high accuracy by combining previously developed probabilistic modeling and calibration methods with parameter optimization based on experimental data (80 experiments involving 5-20 real robots). Key properties of the optimization process are outlined with the help of a simple scenario that can be solved analytically, and validated numerically for the more complex, non-linear inspection scenario.