Will need hands on to truly tell, but the doom demo seems very promising. If it can play that with text descriptions of where stuff is by distance and degrees in a 3D context then many GUI automation tasks should be easily doable