Improving Generalization Ability of Robotic Imitation Learning by Resolving Causal Confusion in Observations