I have played with this and there is a way to do it, but perhaps not too cleanly.
It is possible to keep the SDL view that is created in the PythonActivity around but not use it as the main view for the app. This will basically run a kivy app in the 'background' of your app with java views.
The next step is to create a kivy app that does not rely on any UI components to handle calls from your java source - but it needs to be an instance of kivys App() to keep python running in the background and it can basically hang out in an infinite while loop.
From there you can write a java interface in python and jnius that supports calling your python code from java.
I've tried this with numpy for a similar 'java-to-numpy' experiment and it took a bit of massaging of the data when going back and forth across the jni boundary but it was able to call numpy functions from a java app.
It takes a bit of work but there is light at the end of that tunnel