Virtual agent positioning driven by scene semantics in mixed reality