In this simulation study, we demonstrate that it is feasible to control a robotic arm using fMRI-based brain-computer interface (BCI), which is a noninvasive method that has the advantage of no surgical risk over invasive BCI methods implanting the microelectrode array into the brain. Moreover, fMRI-based BCI has better spatial resolution compared to EEG-based BCI, which is a primarily used noninvasive BCI method. Using fMRI-based BCI, we focus on decoding the trajectories of arm movements in the primary motor cortex (M1). We construct the distribution map of the preferred direction of the motor cortex neurons and model the fMRI signals, which is based on the map adopting the fMRI forward model and generated by direction of arm movements. The result show that directional sensitivity, the key element of the motor decoding, can be found at the voxel level analyzed using multi-voxel patterns in order to compute the directional tuning curves. Furthermore, the results also demonstrate that the trajectories of arm movements can be predicted with six second delay from the estimated directional tuning curves. Therefore, this study provides that if we find the directional tuning properties during motor imagery, the hemodynamic responses measured by fMRI can be used to control the robotic arm between thought and movement.