Continuous vision-and-language navigation (VLN) is usually judged by whether an agent reaches a goal, on the assumption that success reflects the agent's ability to carry out language-grounded behavior. A new arXiv preprint challenges that assumption, arguing that it can fail.
The authors describe the failure as a 'physical executability gap' between measured goal-reaching and actual physical execution of instructions. The title indicates the work both reveals and bridges this gap, though the abstract provided is cut off, so the specific bridging approach is not described in the available text.
This is a single preprint, so there are no independent sources to compare or contrast. The claim matters because it questions whether a widely used evaluation metric in VLN research truly captures what an agent can physically do; if the gap is real, agents may look more capable than they are.