QT-Opt

该数据集由谷歌大脑团队创建，旨在支持基于视觉的机器人抓取技能学习。数据集包含超过 58 万次真实世界的抓取尝试，涵盖约 1000 种视觉和物理特性各异的训练对象。数据来源于 7 台 KUKA LBR IIWA 机器人，配备单目 RGB 相机和两指夹爪。数据收集过程采用自主监督学习，通过机器人自身判断抓取成功与否进行标注。数据集的应用领域主要集中在机器人视觉引导抓取技能的强化学习研究，助力机器人在复杂环境中实现高效、动态的抓取操作。This dataset was developed by the Google Brain team to support vision-based robotic grasping skill learning. It contains over 580,000 real-world grasping attempts, covering approximately 1,000 training objects with distinct visual and physical characteristics. The data was collected using 7 KUKA LBR IIWA robots equipped with monocular RGB cameras and two-finger grippers. The data collection process adopted autonomous supervision, where annotations were generated based on the robot's own judgment of grasp success or failure. The main application areas of this dataset focus on reinforcement learning research for vision-guided robotic grasping skills, helping robots achieve efficient and dynamic grasping operations in complex environments.

搜集汇总

背景与挑战

背景概述

QT-Opt是一个可扩展的自我监督视觉强化学习框架，通过超过58万次真实抓取尝试训练深度神经网络，实现了96%的未见物体抓取成功率，并能自动学习重新抓取策略和动态响应扰动。