I also wonder how accurate their estimates are of arm and leg poses, much less their motions. I do believe they can detect the location of a torso. I don't believe they can do the same for forearms, hands, legs, or feet, or their precise motions.

I suspect the articulated pose and motion of their animated stick figures were extrapolated from progressions of articulations of limb poses that were learned from the light-based training videos, and then they were overlaid onto the wifi-based signal of only the moving trunk.

All that animation does look cool, though.