With further training I don't think it's possible: since those movements are not useful nor harmful, they will appear in winning and losing matches. A possible

With further training I don't think it's possible: since those movements are not useful nor harmful, they will appear in winning and losing matches. A possible solution might be including some distance traveled metric in the reward function...