Efficiently avoiding saddle points with zero order methods: No gradients required