This is a pretty vacuous vox pop study that doesn't really tell me anything I didn't know.
The problem with this approach goes back to BF Skinner and his teaching machines in the 1950s. Essentially it is that all the interaction has to be scripted, and if you think about even the large free roaming games like GTA, all the key interactions are pre-determined.
The problem with humans is that they do not act in linear predictable ways, and that is what makes them so interesting, and challenging. A VR environment can not yet portray the level of detail necessary for complex human-human interaction to be realistic.
The problem with medical students is that progressive generations of well meaning medical education 'innovation' mean that they spend less and less time interacting with patients. Only this, structured and supervised properly, is good training for what you want them to be able to do at the end - to interact with patients.
I do see a time in the future when some good learning will be possible in a true virtual environment, but for now, like other simulation based training, it is limited to the relatively few situations when the situation portrayed is adequately realistic and the stuff being taught is simple e.g. Pavlov's dog stimulus-response stuff - things like resuscitation. It is not appropriate for teaching, even less for testing, complex human-human interaction.
[CoI IAAD, have masters in MedEd, and teach in (allegedly) the top medical school in the UK]