VICtoR: Learning Hierarchical Vision-Instruction Correlation Rewards for Long-horizon Manipulation