Fearful Goal Generation for Reliable Policy Learning