ReVLA: Reverting Visual Domain Limitation of Robotic Foundation Models