From what I've observed, (so preface the following with "It seems to me that:") BD obviously has a well-honed "playback" tool. They have the ability to command the robot to operate in a new environment, record its own movements to sub-millimeter accuracy, and record the environment (it has a live 3d point map after all.). Then, they load in the recorded environment+motion data, and play it back as a 3D scene. Naturally, this would be needed to analyze the performance and make software and hardware iterations. However, this data is likely also used to re-create a digital scene for the purposes of cgi enhancement of a performance that's intended to be recorded and released to Youtube. Similar to how Favreau uses a giant screen to approximate the lighting on the chroma-keyed subjects[1] as they perform, this data -- a giant shiny rectangle goes here, a window and a flashing blue light bulb there, these steps over here -- combined with accurate camera motion data, is then used to create a digital model with the exact shadows that one would need to create a more realistic "fake video" in the first place. The trick is that Boston Dynamics then takes the original take (or iterations of takes of it running the exact same sequence), and iterate a cgi-enhanced version of their "cool dance video" that the marketing team then signs off on. "Ohh yeah, that's what our vision is. We don't want them to see shaky appendages, or micro-stutters. Add in some extra scuff marks, too! Perfect. Upload it."
[1] https://illumin.usc.edu/the-volume-how-the-mandalorian-revol....