I decided to experiment with ml-agents and to make a game where worker type of players push target in to goal and the aggressors type players attack each other. Aggressor learned something quickly but it was expected. Because they were made from examples of ml-agents assets. But workers have different observations and action thus they require much time and effort. The result you can see in second video on top of this post. After a week of my experiments my workers behave better :). And i think that i will finish this project!