Recent vision-language-action (VLA) models have impressed in local robot manipulation, but a new arXiv paper argues that their success does not readily carry over to long-horizon mobile manipulation. The authors report experiments with a pretrained subtask VLA model, though the abstract cuts off before presenting specific results.
The paper's title, "Cracking Mobile Manipulation: From Failure Diagnosis to Agentic Robot System," indicates a two-part approach: first diagnosing why VLA models fail on extended mobile tasks, then using those insights to build an agentic robot system. The abstract frames this as a necessary step beyond relying on subtask-level VLA performance.
Because the abstract is truncated, details about the failure modes or the proposed architecture are not available. The significance lies in the explicit gap identified between local and long-horizon mobile manipulation, and the suggestion that closing it requires more than scaling VLA models alone.