The DLG collector as used by Caml Light does what you describe.
Updating an object in the global heap to point to a thread-local heap can cause unbounded work to be done before that update as the entire pointed-to structure has to be copied into the global heap first. GC designers really hate write barriers (sections of code that are performed before a pointer is updated) that perform unbounded work.
It would solve the problem with "stop the world" GC dynamics and increase concurrent collecting performance while also adding native fibers to the JVM (and thus make it possible to write highly concurrent software in Java much more easy).