SPOC
收藏资源简介:
SPOC(Shortest Path Oracle Clone)是由 Allen Institute for AI 等机构开发的用于机器人导航与操作的数据集。该数据集旨在通过模仿仿真环境中的最短路径规划器,训练机器人在真实世界中高效导航与操作。数据集包含约 20 万套程序化生成的家庭环境,涵盖 41133 个独特 3D 资产,以及数百万帧的专家轨迹数据。这些数据基于 AI2-THOR 模拟器和 ProcTHOR 框架生成,通过高效启发式规划器利用丰富的环境信息,生成高质量的训练轨迹。SPOC 数据集的创建过程充分利用了大规模程序化生成技术,结合了大量 3D 资产和多样化场景布局,确保了数据的多样性和丰富性。其应用领域主要集中在机器人导航、目标定位、物体抓取与搬运等任务,旨在解决机器人在复杂环境中自主决策和操作的问题。该数据集的开放性使其能够为机器人学习领域提供强大的支持,推动机器人在真实世界中的应用。
SPOC (Shortest Path Oracle Clone) is a dataset for robot navigation and manipulation developed by institutions including the Allen Institute for AI. This dataset aims to train robots to perform efficient navigation and manipulation in the real world by mimicking the shortest path planners in simulated environments. The dataset contains approximately 200,000 procedurally generated household environments, covering 41,133 unique 3D assets and millions of frames of expert trajectory data. Generated based on the AI2-THOR simulator and ProcTHOR framework, these data utilize rich environmental information via efficient heuristic planners to produce high-quality training trajectories. The development process of the SPOC dataset makes full use of large-scale procedural generation technology, combining a large number of 3D assets and diverse scene layouts to ensure the diversity and richness of the data. Its application scenarios mainly focus on tasks such as robot navigation, target localization, object grasping and handling, aiming to address the problems of autonomous decision-making and manipulation of robots in complex environments. The openness of this dataset enables it to provide strong support for the field of robotic learning, promoting the real-world applications of robots.




