As diverse robotic hands keep emerging for industrial and household use, designing general grasp synthesis algorithms applicable to multiple grippers remains challenging. To improve the generality and effectiveness of multigripper planar grasping algorithms, we propose a grasping framework featuring gripper-agnostic scene inference and gripper-changeable optimization. In our approach, we introduce an interaction probability map that bridges the scene inference and grasp optimization modules. It efficiently decouples the learning of grasping knowledge and modeling of gripper's kinematics. The inference module adopts a modified directional ensemble method with a generated fingertip dataset to refine scene information. In grasp optimization, we formulate gripper-kinematic constraints for different grippers according to joint types. Extensive evaluations on the Cornell Grasping Dataset (with a success rate of 95.51%) and on multifingered grippers (ten grippers in the real world) demonstrate that our hybrid approach generalizes learnable knowledge across various grippers. This work enables the direct transfer of learned grasping knowledge to new grippers in real-world applications.

中文翻译：

---

跨夹具转移抓取：用于广义平面抓取生成的学习优化混合框架


随着工业和家庭用途的机器人手不断涌现，设计适用于多个夹具的通用抓取合成算法仍然具有挑战性。为了提高多夹具平面抓取算法的通用性和有效性，我们提出了一种具有夹具不可知场景推理和夹具可变优化的抓取框架。在我们的方法中，我们引入了一个交互概率图，它连接了场景推理和掌握优化模块。它有效地解耦了抓取知识的学习和夹具运动学的建模。推理模块采用改进的方向集成方法和生成的指尖数据集来细化场景信息。在抓取优化中，我们根据关节类型为不同的夹具制定夹具运动学约束。对康奈尔抓取数据集（成功率为 95.51%）和多指抓手（现实世界中的十个抓手）的广泛评估表明，我们的混合方法概括了各种抓手的可学习知识。这项工作能够将学到的抓取知识直接转移到现实应用中的新夹具上。