StopWatch: Goal Grounding for Object-goal Navigation

StopWatch. Goal Grounding for Object-goal Navigation. By Xinhao Liu and six others, from Amazon FAR, New York University, and two other institutions. Published on the sixteenth of October. Its code repository is public and real: about thirty source files.

Fine-tuning a pretrained vision language model into a navigation policy has become a common recipe for object-goal navigation. These models bring rich world knowledge and semantic priors to the job. Success rates nonetheless stay low, and it has been unclear which capability the fine-tuned policies are missing. The paper's diagnosis names one: goal grounding is the dominant failure. The policies explore well. Where they fall down is noticing that the goal is already in view, and estimating how far away it is. They search a room and fail to recognize when the search is over.

If that diagnosis is right, grounding the goal explicitly should help, and the authors do exactly that. Explicit grounding substantially improves success and reaches state-of-the-art results on six benchmarks drawn from three scene collections. Along the way, the authors correct annotation defects in those benchmarks, so that future methods are evaluated more faithfully. On the cleaned benchmarks, an analysis supplies the sharpest piece of evidence. Given ground-truth grounding, their policy nearly matches the expert that generated its demonstrations. In other words, once the policy is told correctly where the goal is and how far, almost nothing else stands between it and the behaviour it was trained to imitate.

That result defines what the claim covers. It locates the remaining headroom in perception rather than in policy learning, for fine-tuned policies of this kind, measured on these cleaned benchmarks. It does not say exploration is solved everywhere, only that these policies already explore well and that grounding is where they lose most. The abstract describes the fix only as grounding the goal explicitly, so the claim is about where the failure sits more than about a particular mechanism for fixing it.

For anyone building navigation policies, the pointer is toward perception: seeing the goal and judging its distance, rather than learning to wander more cleverly. The paper positions goal grounding as a missing piece of object-goal navigation, and offers a simple, effective way to supply it.