Are models any good at descerning motion from multiple frames?

For instance if I gave models multiple animations of a bouncing ball as individual frames. Would they be able to tell which bounce was the more realistic motion.

(Is this a potential new benchmark? maybe also variations of stair dismount)