Robonet

RoboNet 数据集由加州大学伯克利分校、斯坦福大学、宾夕法尼亚大学和卡内基梅隆大学联合创建，旨在通过大规模数据共享推动机器人学习的泛化能力。该数据集包含来自 7 种不同机器人平台的约 16.2 万条轨迹，涵盖近 1500 万帧视频，涉及数百个物体、多种视角和环境。数据集以自监督方式收集，无需人工干预，且易于扩展至新的机器人硬件和传感器。其创建过程通过跨机构合作，将不同实验室的机器人交互数据整合，形成多样化数据池。RoboNet 的应用领域主要集中在视觉操控任务的泛化研究，例如物体重定位、抓取与放置等，旨在解决机器人在新场景、新物体、新机器人平台上的快速适应问题。

The RoboNet dataset was jointly developed by the University of California, Berkeley, Stanford University, the University of Pennsylvania, and Carnegie Mellon University, with the goal of advancing the generalization of robotic learning through large-scale data sharing. This dataset contains approximately 162,000 trajectories from 7 distinct robotic platforms, covering nearly 15 million video frames, and involves hundreds of objects, multiple viewpoints, and diverse environments. Collected in a self-supervised manner without human intervention, the dataset is easily scalable to new robotic hardware and sensors. Created via cross-institutional collaboration, it integrates robotic interaction data from various laboratories to build a diverse data repository. The primary application areas of RoboNet focus on generalization research for visual manipulation tasks, such as object relocalization, grasping and placing, aiming to address the rapid adaptation challenges of robots in novel scenarios, new objects, and unfamiliar robotic platforms.

搜集汇总

背景与挑战

背景概述

RoboNet是一个开放的机器人交互视频数据库，包含来自7种不同机器人平台的1500万帧视频数据，旨在支持视觉机器人操控的通用模型研究。该数据集由4个机构共同贡献，具有多样性和大规模特点，且完全开源。