The obvious would be to exclude the NaN observations but you may run into an issue that the number of replicates for each group must be the same so you'll also have to remove a matching replicate for each factor combination matching the missing one(s). The result is you may not have any data left for a given set of factors; all that would depend on the design and how many and where the NaN values are located.
IOW, it's quite possible (maybe probable?) that you simply can't apply the basic Friedman test to the dataset.
The <Skillings Mack test> described at the link is a Friedman-like test that can handle missing data but a quick search didn't find a ready MATLAB implementation and I haven't built one, either.
