Then you would have to deal with the effect of subjecting the mice to such a degree of control, precision of action, predictability of environment, and other factors. As was found with the 'Rat Park' experiments in the 70s, experiments on rats in cages are not informative about the behavior of rats not in cages. It is tremendously difficult to study systems as complex as living animals, and we ignore "unimportant" factors at our peril. How different would our society be if the Rat Park experiments, which showed that rats addicted to cocaine will naturally ween themselves off of the drug when provided with a stimulating social environment, was reported on first instead of the one we heard about instead that claimed cocaine was so addictive that a rat which takes a dose becomes inescapably addicted and will dose themselves with it until they die? (Which they will.... if they are locked in a cage and the alternative is just living in a miserable cage.)

Rat Park can only sometimes be replicated:

https://www.ncbi.nlm.nih.gov/pubmed/9148292?dopt=Abstract

https://www.ncbi.nlm.nih.gov/pubmed/2616610

https://www.ncbi.nlm.nih.gov/pubmed/3696469

https://www.ncbi.nlm.nih.gov/pubmed/18463628

SSC wrote an article about it:

http://slatestarcodex.com/2017/10/25/against-rat-park/

Even automation would not solve this issue. If these results are scent related, then it is quite possible that different automation methods would produce different scents, or even the same method, in a different environment / location.

This is a fascinating issue.

But you could at least use the same automation methods across different study groups within the same study, eliminating potential confounding variables between your groups.