Title: GIFT: Geometry-Induced Functional Transfer for Category-level Object Manipulation

Abstract:

Executing robotic manipulation tasks with unfamiliar objects in novel settings remains difficult, primarily due to constraints in generalization. To overcome this, we introduce GIFT (Geometry-Induced Functional Transfer), a novel skill transfer framework that allows robots to acquire complex manipulation skills and constraints from a single human demonstration. Our methodology tackles the dual challenges of learning and performing tasks by extracting geometric representations that emphasize interactions centered on the object. By utilizing the Functional Maps (FMC) framework, we efficiently align interaction functions between objects and their surroundings. This capability enables the robot to replicate operations across items sharing similar topologies or categories, even if their physical shapes vary drastically. Furthermore, we integrate screw interpolation (ScLERP) to synthesize smooth, geometrically informed robot trajectories, ensuring that transferred skills strictly comply with the original task constraints. We confirm the robustness and adaptability of our approach through comprehensive experiments, which show successful skill transfer and task execution across various real-world scenarios without the need for further training.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC