Did you ever figure out how to achieve this?
I intent to display stereoscopic 360° images. The basic idea is to show 2 equirectangular images on 2 UV-spheres, position 2 cameras inside the spheres, then ignore the HMD position and only use rotation to point the cameras in sync.
A sphere could be parented to the ARVRCamera node as a workaround for keeping it centered, but that still doesn't solve the problem of targeting both eyes separately.