Tabletop Object Manipulation Dataset

该数据集是通过人机交互收集的，涵盖了106种不同的对象类型，用于训练和测试操作技能。该数据集包含了多种操作任务，如“捡起”、“移动到附近”、“敲击”、“直立放置”和“放入”等，这些任务在多个评估场景中展开，同时场景中还随机放置了干扰物。在规模上，该数据集为训练提供了106种不同的对象类型，并为评估设置了13种已见对象和10种未见对象。该数据集的任务旨在研究开放世界环境中的对象操作和泛化能力。

This dataset was collected via human-computer interaction, covers 106 distinct object types, and is dedicated to training and testing manipulation skills. It includes various manipulation tasks such as pick up, move to nearby positions, tap, place upright, and put into. These tasks are conducted across multiple evaluation scenarios, where distractors are randomly placed in the environments. In terms of scale, the dataset provides 106 distinct object types for training, and reserves 13 seen objects and 10 unseen objects for evaluation. The tasks of this dataset are designed to study object manipulation and generalization capabilities in open-world environments.

搜集汇总

背景与挑战

背景概述

该数据集是一个通过人机交互收集的大规模操作数据集，涵盖106种对象类型，包含多种操作任务如捡起、移动等，并在评估中设置干扰物以模拟开放世界环境。其特点在于强调泛化能力，通过已见和未见对象的区分来测试模型在真实场景中的适应性。