Future-guided offline imitation learning for long action sequences via video interpolation and future-trajectory prediction.