Researchers led by Zequan Zeng have presented MARS (Modular Atomic-skill Routing System), which combines a general vision-language-action model with specialised skills learned by reinforcement learning for manipulation tasks with several stages, according to a paper on arXiv. General models still lack precision and stability at critical stages such as grasping, aligning and inserting objects, the authors write.
MARS keeps the general model unchanged and trains separate policies for those critical stages. A router that looks at the recent course of the task and at descriptions of each skill decides which policy is in control at each moment.
On four LIBERO task suites, the router's choices reached an average Macro-F1 score of 90.47%, 7.08 percentage points above a simpler router. In a simulated tube transport task, full-task success rose from 20.04% to 85.67%.



