I'm having this problem myself and I don't think a solution currently exists to do this in the way you're describing.
The gg2list function contained in the call to ggplotly expects to be able to iterate over a ggplot object's layers to create the corresponding plotly object. If you step into the ggplotly function with RStudio's debugger, you can see the various ways in which it attempts to iterate across the object it receives in order to extract its properties.
It doesn't know what to do with the object returned by the arrangeGrob function because it's not just a single ggplot object (it's a grob arrangement, etc).
Calling class() on the objects in question illustrates this point somewhat.
> class(g1)
[1] "gg" "ggplot"
> class(g)
[1] "arrange" "ggplot" "gTree" "grob" "gDesc"
I think in order to have multiple plots heads up in the same plotly object, we will need to use the facet options in ggplot or the native plotly R bindings. Unfortunate, because gridExtra is very powerful and flexible, but the ggplot translation mechanism doesn't seem to be able to handle it.