Vision-language models are increasingly used to give mobile robots goals stated in ordinary language. This survey covers 118 papers published between 2021 and 2026 and groups them by where the model sits in the stack: as a perception front end, as a planner, or as an end-to-end policy. For each group we summarise the benchmarks used, the assumptions made about the environment, and the reported failure modes. We find that reported success rates are not comparable across papers, because more than half of the works define their own episode termination rule, and we propose a minimal reporting checklist that would make them comparable.