I think attaching the camera to the object makes it only harder to predict the behavior. Manually updating camera position and rotation with the robot movement sounds easier.
Take a look at this example. If I understood the problem correctly, this is what you are looking for.