Whereas robust processing techniques are routinely used for estimating high-quality magnetotelluric (MT) transfer functions, such techniques are not commonly applied for controlled-source electromagnetic (CSEM) processing, although CSEM and MT data suffer from similar noise. We implemented a new CSEM processing scheme that combines CSEM-specific preprocessing with statistically robust least-squares stacking to extract interpretable ground responses from very noisy onshore CSEM data. We applied the robust processing to signals from a new CSEM transmitter that was equipped with three grounded electrodes and allowed us to generate signals at multiple source polarizations with relatively little field effort. For this transmitter setup, we formulated a bivariate relation between the source currents injected through any two of the three source electrodes and the recorded electromagnetic field components. The resulting weighted least-squares system of equations from which we determined ground responses allowed us to jointly process data from multiple polarizations. Using several polarizations resulted in more stable response estimates than can be obtained from standard configurations with two distinct source dipole orientations. Exploiting dependencies between the three basic response functions that we obtain allows consistency checking and demonstrates the stability of our robust processing scheme. From the basic responses, data at arbitrary source polarizations can be synthesized, which may be useful for optimizing target illumination. We tested the benefits of robust CSEM processing using examples of data recorded across the CO2 storage test site at Ketzin, Germany, in an area heavily affected by various sources of strong cultural noise, including impressed-current cathodic protection systems, wind power plants, and major power lines.