The abstract of a new arXiv paper opens by noting that embodied AI has not enjoyed the same data-rich pretraining as language models. As a result, adding a new manipulation capability typically means paying for demonstrations on every task, or running a frontier model in every episode. The authors say they are studying a third route, but the provided text cuts off before the route is fully described.
The paper's title points to the idea: "Compile Once, Run Cheap: Frozen, Portable Skills for Embodied Agents, and What Recursive Self-Improvement Adds." This suggests a method where skills are compiled once and then reused cheaply, rather than being re-derived per task or per episode. The abstract also signals that the authors are examining whether recursive self-improvement can build on this approach.
Because the abstract is truncated, the exact mechanism of the third route is not available in the source. The significance, as stated, is that a middle path may exist between expensive per-task data collection and expensive per-episode inference, potentially making embodied manipulation more scalable.