Grounded worlds and the planning gap
Predicting the next frame and planning toward a goal look adjacent. They are not, and the second does not fall out of the first.
Work on grounded world models for semantically generalizable planning targets exactly that gap.
Two different capabilities
A model that accurately forecasts what happens next has learned dynamics. A model that can plan has learned dynamics plus a way to search over them toward an objective. Adding the search is not a matter of scaling the prediction.
Semantic generalisation is the harder half. Planning inside a known task distribution is well studied. Planning toward a goal described in language the training set never contained requires the world model and the language grounding to share a representation, which most architectures do not arrange.
Why it reaches past robotics
It is the same architecture question facing any agent operating over long horizons in a partially observed environment. The robot case just makes the failure visible, because a bad plan knocks something over instead of producing a plausible paragraph.
And it connects to the data question. Web data supplies perception and semantics; robot data supplies dynamics and control. A training scheme that routes each to the parameters it should shape is doing something more structured than blending, and grounding is the word for getting that routing right.
The measurement problem
Planning benchmarks are notoriously easy to overfit, because a task distribution rich enough to be interesting is usually narrow enough to memorise. Any claim about generalisable planning needs held-out goals that the authors did not design, and very few papers in this area have them.
That is not a criticism of this work specifically. It is the standing weakness of the subfield, and it is worth stating whenever a planning result is reported.
arXiv — Grounded world model for semantically generalizable planning → · Hugging Face Papers — Agents' Last Exam → · arXiv — LLM reasoning is latent, not the chain of thought →