Predicting whether terrain is traversable is a core challenge for autonomous robots, but the datasets used to train such predictors are often generated by physically deploying robots in the field. That process is expensive and may not cover the variety of environments a robot will actually encounter. A new preprint on arXiv introduces ZeST, a method that instead relies on vision-language models (VLMs) to assess traversability in a zero-shot manner, meaning it does not require task-specific training data.

The paper argues that VLMs, which already encode broad world knowledge, can be adapted to reason about terrain from visual input alone. By framing traversability as a language-grounded problem, ZeST aims to generalize to unknown environments without the need for robot-collected examples. The authors position this as an alternative to traditional dataset-generation pipelines, which often involve putting robots through extensive data collection.

Because the abstract is truncated, the full details of the architecture and any experimental results are not available in the source. The announcement does not include comparisons to other methods, so the claims should be read as the authors' own framing. The work is presented as a replacement version of the paper, indicating it has been revised since its initial posting.