Seeing Through Pixel Motion: Learning Obstacle Avoidance from Optical Flow with One Camera