Move to See Better: Self-Improving Embodied Object Detection