I have worked in computer vision for two decades. Short answer: Yes, they edit out the arm through clever calibration and matching algorithms. This is a completely automatic procedure (if you have the right software, which I am sure they have).

The image you see is a panorama stitched together out of at least a handful of images. The parts blocked by the arm in one image will be visible in other images. Since you know the camera pose and the arm pose (and the cameras are calibrated as well), you can model exactly what pixels you can use and which need to be replaced. Given at least 6 matching points in the image, you can additionally solve for the camera pose between images and fuse this information with the arm pose to increase accuracy.

You can do similarly nice things by taking a sequence of images from an object or in your house or backyard and feeding these images into a program called "meshroom". It solves for camera position and orientation (6DOF) and then stitches together a point-cloud and can even mesh and texture it for you. This process is called "structure from motion" and it is heavily used in 3D modelling for movies and games to recreate objects without having to tediously model by hand.