GREAT: Geometry-Intention Collaborative Inference for Open-Vocabulary 3D Object Affordance Grounding

Recently, most existing methods [2, 6, 18] establish explicit mappings between semantic affordance categories Open-Vocabulary 3D object affordance grounding aims and geometric structures, restricted to predefined seen categories to anticipate "action possibilities" regions on 3D objects and fail to ground object affordance out of the training with arbitrary instructions, which is crucial for robots to categories. Thus, some studies [27, 42, 50, 56, 58] generically perceive real scenarios and respond to operational explore grounding object affordance through additional instructions, changes. Existing methods focus on combining images encompassing combining images or languages or languages that depict interactions with 3D geometries that depict interactions with 3D geometries to introduce to introduce external interaction priors. However, they external interaction priors, and mitigate the generalization are still vulnerable to a limited semantic space by failing to gap lead by affordance diversity. Despite their remarkable leverage implied invariant geometries and potential interaction progress, they are still vulnerable to a limited semantic intentions. Normally, humans address complex tasks space by failing to leverage implied invariant geometries through multi-step reasoning and respond to diverse situations among objects with the same affordance, as well as potential by leveraging associative and analogical thinking.