CodeAuditor

VLMs Cannot Plan, But Can They Formalise?

Vision Language Models Cannot Plan, but Can They Formalize?

Read the original source

What the paper says

The paper argues that vision-language models struggle with direct long-horizon planning, but can work better as formalizers: translating visual planning problems into formal representations such as PDDL for a classical planner.

The Critique

This is a practical hybrid approach, but it shifts the hard problem into representation. A planner can only solve the world it is given. If the VLM misses an object relation, misreads a scene, or omits a constraint, the formal plan can be logically valid and physically wrong. The system may look more rigorous because it uses a formal solver, while the fragile part remains the visual translation.

Why It Matters

Hybrid AI systems are probably the near-term route to reliable embodied planning. But formal methods do not rescue a bad world model.

What They Missed

End-to-end error propagation, noisy real-world scenes, uncertain object states, and recovery mechanisms when the formalisation is incomplete.

The Big Question

Is the VLM a reliable bridge to symbolic planning, or the weakest link disguised as a translator?