BC-Z

BC-Z 数据集由 Robotics at Google、Everyday Robots、UC Berkeley 和 Stanford University 等机构联合创建，旨在助力机器人通过模仿学习实现对新任务的零样本泛化。该数据集包含 25,877 个机器人操作演示片段，涵盖 100 种不同的视觉操作任务，如抓取、放置、擦拭等，总时长超过 125 小时。数据来源于 12 台机器人，由 7 位操作员通过虚拟现实遥操作设备收集，同时采集了 18,726 段人类执行任务的视频。数据集的创建过程结合了专家演示和人类在环的共享自主干预，以纠正机器人策略错误，提升数据质量。BC-Z 数据集主要用于机器人视觉操作任务的模仿学习研究，致力于解决机器人在无额外数据情况下对新任务的快速适应问题，推动机器人在复杂环境中的通用性发展。

The BC-Z dataset was co-created by institutions including Robotics at Google, Everyday Robots, UC Berkeley, and Stanford University, aiming to empower robots to achieve zero-shot generalization to novel tasks via imitation learning. This dataset contains 25,877 robotic manipulation demonstration clips, covering 100 distinct visual manipulation tasks such as grasping, placing, wiping, etc., with a total duration exceeding 125 hours. The data was collected from 12 robots by 7 operators using virtual reality teleoperation equipment, and 18,726 videos of humans performing the tasks were also captured simultaneously. The dataset construction process integrates expert demonstrations and human-in-the-loop shared autonomous intervention to correct robotic policy errors and enhance data quality. The BC-Z dataset is primarily utilized for imitation learning research on robotic visual manipulation tasks, and aims to address the challenge of rapid adaptation of robots to novel tasks without additional data, thereby advancing the development of robot versatility in complex environments.

搜集汇总

背景与挑战

背景概述

BC-Z是一个用于机器人模仿学习的大规模数据集，包含100个操作任务的VR远程操作演示，支持通过自然语言或视频指令进行任务泛化。该数据集的特点是通过多样化的任务训练和条件化信息（如自然语言嵌入或人类视频）实现零样本任务泛化，适用于复杂的长期任务场景。