●builderYou can use this framework to extend deployed VLA models to new manipulation tasks without collecting fresh human demonstrations for every skill gap.
●researcherThe two-stage pipeline — VLM-guided segmentation plus autonomous primitive discovery — offers a concrete architecture for self-improving robot policies without human re-annotation.