VLA Handover lets robots call a large server model only when a small one may fail

Takayuki Nishio has described VLA Handover in an arXiv paper, a way for a robot to run a small vision-language-action model on board and switch to a large one on a server only when the small model is likely to fail. Large models handle new situations better but are hard to run in real time on a robot, and asking a server for every action limits how many robots one server can support.

The method treats switching as a cost problem with a required success probability and uses a heuristic that needs no retraining of either model. Tested in simulation on three of the four LIBERO suites, using SmolVLA as the small model and pi-0.5 as the large one, it raised success by 11 to 22 percentage points over the small model alone, and sent 49 to 77% fewer requests than server-only control at a success rate 2 to 4 points lower.