In this paper, we presents a novel non-invasive brain-Geminoid control system by using single-trial classification of bimanual movements to achieving a noninvasive brain-computer interface (BCI) with many control dimensions and easily interact with an outdoor complex environment in real-time. Two BCI-naive subjects performed or imagined performing 4 movements of bimanual hand during the measurement of magnetic fields to control a Geminoid HI-2 (humanoid robot) through a multidimensional BCI. We applied a nonlinear support vector machine (SVM) to classify the 4 bimanual hand movements using 114 magnetoencephalography (MEG) sensors over the sensorimotor cortex. The mean classification accuracy of a 2-class decoding was suitable for real-time brain-Geminoid control application, with classification accuracies equivalent to those obtained in precedent BCI studies involving uni-manual movements. Moreover, our results demonstrated that decoding bimanual hand movements in real-time using the amplitudes of the event-related magnetic fields is very promising to implement multidimensional-control based BCIs.