We propose GRACE, a training-free framework that bridges VLM reasoning and precise robot execution through executable analytic concepts, enabling semantic-to-physical grounding and zero-shot generalization for articulated-object manipulation tasks.

We propose the Articulated Object Procedural Generation (Arti‑PG) toolbox, which synthesizes diverse and unlimited articulated objects with rich annotations through program‑oriented structure manipulation and analytic label alignment, addressing the scarcity of high‑quality articulated object data and annotations.

We present AOT and the AOT-driven baseline AOTNet to empower machine intelligence with the kind of human capability to understand and then interact with articulated objects, espeically the novel ones, at the conceptual level for the first time.

We present ConceptFactory, a novel scope to facilitate more efficient annotation of 3D object knowledge through object conceptualization, containing a unified toolbox for object conceptualization and a large collection of conceptualized objects.