PhysVLM: Enabling Visual Language Models to Understand Robotic Physical Reachability