Mental rotation is the ability to mentally rotate two-dimensional and three-dimensional objects in one's mind. Previous studies have examined the relationship between mental rotation abilities and embodied cognition. However, conflicting findings underscore the necessity for further investigation. This study aims to investigate how different levels of embodiment in mental rotation training affect mental rotation training. Three versions of mental rotation training with different levels of embodiment (keyboard, touchpad and augmented reality) were developed for this study. A quasi-experimental study was conducted with a total of 125 university students to explore how different training versions affect participants' training performances, perception of the usability of the training and mental rotation ability. The results suggest those who used the versions with higher levels of bodily engagement performed better than those who used the versions with lower levels of bodily engagement (p < 0.001). Participants' mental rotation ability also improved significantly after training with the augmented reality (AR1) version compared to those in the keyboard group (p = 0.01). The learning analytics results provide additional clarity by indicating that the participants' mental rotation ability can be predicted by their training performance in the AR condition (p < 0.001). This study has significant implications for the potential of using embodied AR training to improve mental rotation ability.