I am planning to use textual rules like traffic rules in motion planning of autonomous car. I can think of using BERT like models to generate embeddings and then use these embeddings for motion or trajectory planning. My question is should I go for things like ontology, knowledge graph instead of embeddings of Natural language processing (NLP) model? Are embeddings of NLP models usable for motion or trajectory planning, or they need some other extra layer after embedding generation?
Note: I have worked with computer vision, where I have used embeddings or feature vector in lot of problem domain, but new in the domain of NLP and knowledge representation.