Answers are grounded in this paper and cite their sources.
Robotic grasping models trained on large‐scale datasets perform well on common objects but often fail on out‐of‐distribution (OOD) objects or nonstandard grippers because of dataset bias and poor physical grounding. To address this, an efficient framework is proposed for fine‐tuning grasping models on novel objects. High‐quality 3D object meshes are reconstructed using Neural Radiance Fields (NeRF), grasp candidates are generated rapidly by a sampling‐based scheme that sweeps a fixed set of gripper orientations over the reconstructed surface and scores them by force closure, and physics‐based dynamic filtering discards unstable grasps. In a randomized simulation benchmark of 25 objects, including six held‐out objects, fine‐tuning with the proposed labels raises the top‐1 grasp success rate from 62% to 91% (2500 trials per method). In real‐robot experiments on nine objects (135 trials per method), the success rate increases from 29.6% to 73.3%. Generating 6500 grasp labels takes about 2.2 h, compared with about 13 h for an antipodal sampling pipeline. The framework provides a scalable label generation pipeline and a physics‐grounded filtering mechanism that improves real‐world grasping robustness. The reliability of simulation‐based filtering, the sim‐to‐real gap, and the simulation setup effort are discussed as limitations.
No takes yet. Share an insight, caveat, or question.
Hu et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: