arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2310.18569cs.RO
收藏

通过改进的微调过程提升新物体的抓取性能

Enhancing Grasping Performance of Novel Objects through an Improved Fine-Tuning Process

Shenzhen University(深圳大学);Xpeng Robotics(小鹏机器人);Chinese Academy of Engineering(中国工程院)

Xiao Hu, Xiangsheng Chen

AI总结:

本文提出一种通过采样抓取方向快速生成新物体6D位姿标注数据的方法,并设计考虑夹爪闭合影响的优化标注框架,使新物体微调效率提升400%。

摘要:

抓取算法已从平面深度抓取发展到利用点云信息,使其能够应用于更广泛的场景。然而,基于在基础开源数据集上训练的模型所得到的数据驱动抓取,在新物体上可能表现不佳,而不同场景中往往需要抓取新物体,因此必须使用新物体进行微调。驱动这些算法的数据本质上对应于手部在6D位姿下的闭合区域,并且由于6D位姿的独特性,通常采用合成标注或真机标注方法。通过真机标注获取大量数据十分困难,因此合成标注成为一种常见做法。然而,使用传统方法获取带标注的6D位姿数据极为耗时。为此,我们提出了一种快速获取新物体数据的方法,从而实现更高效的微调。我们的方法主要对抓取方向进行采样,以生成并标注抓取。实验结果表明,针对新物体的微调过程比其他方法快400%。此外,我们提出了一种优化的抓取标注框架,该框架考虑了夹爪闭合的影响,使标注更加合理。本文被接收后,我们将以开源形式发布我们的算法。

英文摘要:

Grasping algorithms have evolved from planar depth grasping to utilizing point cloud information, allowing for application in a wider range of scenarios. However, data-driven grasps based on models trained on basic open-source datasets may not perform well on novel objects, which are often required in different scenarios, necessitating fine-tuning using new objects. The data driving these algorithms essentially corresponds to the closing region of the hand in 6D pose, and due to the uniqueness of 6D pose, synthetic annotation or real-machine annotation methods are typically employed. Acquiring large amounts of data with real-machine annotation is challenging, making synthetic annotation a common practice. However, obtaining annotated 6D pose data using conventional methods is extremely time-consuming. Therefore, we propose a method to quickly acquire data for novel objects, enabling more efficient fine-tuning. Our method primarily samples grasp orientations to generate and annotate grasps. Experimental results demonstrate that our fine-tuning process for a new object is 400 \% faster than other methods. Furthermore, we propose an optimized grasp annotation framework that accounts for the effects of the gripper closing, making the annotations more reasonable. Upon acceptance of this paper, we will release our algorithm as open-source.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑