Embodiment is expected to play a key role in learning in VR. This randomized control trial determined how the two factors of embodiment (low versus high) and platform (2D PC versus 3D Virtual Reality (VR)) affected STEM learning. A total of 214 undergraduate participants were randomly assigned to four conditions in a 2 x 2 design with pretests and posttests. The low embodied groups did not control the virtual net in the natural selection game, they observed a playback of butterflies being captured. All groups showed significant content knowledge gains by posttest. There was a main effect of embodiment; the active groups that controlled the virtual net learned more content. There was not a main effect of platform on learning. However, there was an intriguing interaction of embodiment by platform by test time revealing that the low embodied VR group learned the least (ES=.38), while the high embodied VR group learned the most (ES=1.07). The low embodied VR players may have expected to have more agency and control over the game mechanics, and playing in the observational mode was infelicitous for learning. Interestingly, the PC condition was not significantly affected when in the more passive/observational mode. Players in the 2D PC condition may be more accustomed to low embodied learning, e.g., watching playback videos on monitors. Several guidelines are recommended for learning design in immersive VR. Non-interactive objects and videos in VR should be used judiciously and designers are encouraged to make the 3D VR content manipulatable and interactive.