At a really high level...

Create an entirely different model. One that acts as a hybrid between a full blown inefficient co-ordinate system and, plausibly, a more efficient box model.

We could start by modelling 'movement' which equates to objects having 1. A reference point, 2. A magnitude and 3. A direction.

It would be beneficial to do this in such a way that one could apply functions to said model achieving a reactive design that gives you a sense of a full blown co-ordinate system while achieving simpler mechanisms for ease in development and efficiency.

I know that's pretty high level stuff, but in my mind I can see a fairly clear path to achieve a significant improvement.