Every sensor is imperfect in its own way: one is precise but slow, another fast but noisy, a third blind in certain conditions. Sensor fusion combines their readings so the weaknesses of each are covered by the strengths of the others, yielding a better estimate than any single sensor gives.
The method weighs each source by how trustworthy it is in the moment, leaning on vision in good light and on other senses when it fades. This blended picture is what lets a robot act confidently despite no single sensor being wholly reliable.