The arXiv preprint ScaleNav tackles language-conditioned local navigation: a robot sees its starting scene, receives a text instruction describing a target location, and must navigate to that spot and stop. The authors frame this as an end-to-end policy learning problem, noting that such policies require diverse training data.
As the title suggests, the proposed solution is to pretrain on internet video at scale. The abstract, however, is truncated in the source, so the specific architecture and experimental results are not described in the available text.
What can be stated from the source is the task definition and the motivation for large-scale pretraining. No comparisons to prior methods or quantitative claims are present in the excerpt, so this digest is limited to the paper's stated scope and approach.