For the next Unicode emoji extension, we might as well just take a word vector model and PCA project it into the Unicode 1 dimensional space. Somebody who is better at linear algebra should comment and tell me how wrong I am.
There's a symbol for the torso, for the fist, for touch and for the strongness of touch. If you hit your own shoulder with a fist with the knuckles facing upwards, you rotate the symbol for the fist, paint the torso and the fist next to each other and put the touch symbol approximately where the shoulder would be.
Even with less granularity, that's still like 256 bytes per grapheme.
I think this would be great for machine translation though.