maze navigation from 3×3 local observations