Answers are grounded in this paper and cite their sources.
Purpose This study aims to address the limited generalization of learned dexterous grasping policies to novel objects. To this end, a general framework is developed in which grasp-part reasoning acts as an intermediate mechanism and primitive policy library reuse serves as the core, enabling effective matching and transfer between novel objects and existing grasping policies. Design/methodology/approach A primitive policy library is first established through continual learning. Then, a large model is used to perform structural part analysis, relevant part identification and grasp-part reasoning on novel objects so as to determine graspable parts that can be matched to the policy library. Finally, dexterous grasping of novel objects is achieved by reusing the corresponding primitive policies. Findings Experimental results validate the effectiveness of the proposed framework. The high-level reasoning process enables more accurate localization of graspable parts that are matchable to the primitive policy library, thereby improving policy reuse quality. Compared with heuristic baselines and methods without part reasoning, the proposed framework achieves superior dexterous grasp success, stability and generalization on novel objects. Originality/value The originality of this work lies in the development of a generalizable dexterous grasping framework that combines grasp-part reasoning with primitive policy library reuse. Instead of using large models for direct low-level control generation, the proposed method uses high-level reasoning to identify reusable graspable parts of novel objects, thereby supporting the transfer and reuse of existing primitive grasping policies. This study provides a new perspective for policy generalization in dexterous grasping.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: