For a non-neural AI take on text-to-scene generation (statistical parsing, symbolic representation, rule-based, 3d models with inverse kinematics), check out Bob Coyne's WordsEye:

http://www1.cs.columbia.edu/~coyne/papers/wordseye_siggraph....

https://www.wordseye.com/

http://www1.cs.columbia.edu/~coyne/