RECON

RECON 数据集由加州大学伯克利分校和卡内基梅隆大学的研究团队创建，旨在支持机器人在开放世界环境中的自主探索和导航研究。该数据集包含超过 5000 条自监督轨迹，涵盖 9 种不同的真实世界环境，如停车场、郊区住宅区、人行道和食堂等。数据集内容丰富，包括机器人在各种环境中的交互数据，如碰撞、颠簸事件以及位置信息等。数据来源为机器人在不同环境中通过时间相关随机行走收集的轨迹数据，这些数据展示了显著的外观变化，包括季节和光照条件的变化。数据集的创建过程基于自监督学习方法，通过检测碰撞和颠簸事件来生成事件标签，从而实现数据的自动标注。该数据集的应用领域主要集中在机器人导航和探索任务，旨在提高机器人在复杂、动态环境中的导航能力，特别是在数据量有限的情况下实现快速适应和目标发现。

The RECON dataset was developed by research teams from the University of California, Berkeley and Carnegie Mellon University to support research on autonomous exploration and navigation of robots in open-world environments. This dataset contains over 5,000 self-supervised trajectories, covering 9 distinct real-world environments including parking lots, suburban residential areas, sidewalks, and cafeterias, among others. The dataset is rich in content, encompassing robot interaction data across various environments, such as collision and bump events as well as location information. The data is sourced from trajectory data collected through temporally correlated random walks performed by robots in different environments, and these datasets exhibit significant appearance variations including changes in seasons and lighting conditions. The dataset creation process is based on self-supervised learning methods, where event labels are generated by detecting collision and bump events to achieve automatic data annotation. The main application fields of this dataset focus on robot navigation and exploration tasks, aiming to improve the navigation capabilities of robots in complex and dynamic environments, especially to enable rapid adaptation and target discovery under the condition of limited data volume.

搜集汇总

背景与挑战

背景概述

RECON数据集专注于机器人自主探索和导航技术，特别是在开放世界环境中的应用。其核心特点是使用学习的潜在变量模型和非参数拓扑记忆，使机器人能够在多样化和未知环境中有效探索和导航。数据集还展示了在未见过的障碍和天气条件下的鲁棒性表现。