In this paper, the problem of forecasting/predicting emergent behavior of untethered magnetic microrobots (MMs) from raw grayscale time-series image data is considered. An interpretable machine learning (ML) method is proposed to forecast/predict such behavior without measuring position, velocity, and/or orientation of each untethered MM. The approach is reduced-order and totally data-driven without knowing any knowledge about the physics and/or interacting dynamics of MMs. The proposed method may provide information/hypothesis about the unknown underlying dynamics of the phenomenon. The method is a promising approach for crossing the reality gap in microrobotics. Experimental results are given in order to validate the proposed method and its effectiveness.