Bridge Data
收藏资源简介:
Bridge Data是一个大规模的多领域和多任务数据集,由加州大学伯克利分校创建,旨在通过跨领域数据提升机器人技能的泛化能力。该数据集包含7200个演示,涵盖71个不同任务,分布在10个环境中,主要围绕家庭厨房任务。数据收集使用低成本的WidowX250s机器人和Oculus Quest2 VR设备。创建过程涉及在多个“玩具”厨房中随机化厨房位置和相机位置。该数据集适用于模仿学习,也可用于离线RL和其他算法。其应用领域旨在解决机器人学习中数据收集成本高和泛化能力有限的问题,通过共享数据集,不同实验室和研究人员可以从中受益,加速机器人技能的学习和泛化。
Bridge Data is a large-scale multi-domain and multi-task dataset developed by the University of California, Berkeley, with the goal of enhancing the generalization capability of robotic skills through cross-domain data. This dataset comprises 7200 robotic demonstrations, covering 71 distinct tasks across 10 environments, with a primary focus on household kitchen-related tasks. Data collection was carried out using low-cost WidowX250s robots and Oculus Quest 2 VR devices. The dataset creation process involved randomizing the positions of both the kitchen setups and the cameras across multiple "toy" kitchens. This dataset is suitable for imitation learning, and can also be employed for offline RL and other relevant algorithms. Its targeted application scenarios aim to resolve the issues of high data collection costs and limited generalization ability in robotic learning. By sharing this dataset, different laboratories and researchers can benefit from it, thereby accelerating the learning and generalization of robotic skills.




