PAC is a benchmark designed to test whether pre‑trained vision‑language models truly understand the object properties, affordances, and real‑world constraints required for executable robot manipulation.