A Billion Ways to Grasp

机器人抓取通常被表述为一个学习问题。随着物理模拟速度和质量的提高，生成用于馈送学习算法的大规模抓取数据集变得越来越流行。一个经常被忽视的问题是如何生成构成这些数据集的抓取。在本文中，我们回顾、分类和比较了不同的抓取采样策略。我们的评估基于 SE(3) 的细粒度离散化，并使用基于物理的模拟来评估相应平行爪抓握的质量和稳健性。具体来说，我们认为 YCB 数据集中的 21 个对象中的每一个对象都有超过 10 亿次抓取。这个密集的数据集让我们可以评估现有的抽样方案 w.r.t。他们的偏见和效率。我们的实验表明，一些流行的采样方案包含明显的偏差，并没有涵盖所有可能的对象可以被抓住的方式。

Robot grasping is often formulated as a learning problem. With the improvement of the speed and quality of physics simulations, generating large-scale grasping datasets for feeding into learning algorithms has become increasingly prevalent. A frequently overlooked question is how to generate the grasps that constitute these datasets. In this paper, we review, categorize, and compare different grasp sampling strategies. Our evaluation is based on the fine-grained discretization of SE(3), and uses physics-based simulations to assess the quality and robustness of the corresponding parallel-jaw grasps. Specifically, we generate over one billion grasps for each of the 21 objects in the YCB dataset. This dense dataset allows us to evaluate existing sampling schemes w.r.t. their bias and efficiency. Our experiments show that some popular sampling schemes contain significant biases and fail to cover all possible ways in which an object can be grasped.

搜集汇总

背景与挑战

背景概述

该数据集专注于机器人抓取学习，通过物理模拟生成了超过10亿次抓取，覆盖YCB数据集中的21个对象，旨在评估不同抓取采样方案的偏差和效率。