DOZE

DOZE是一个专为动态环境中开放词汇零样本物体导航设计的数据集，由北京大学创建。该数据集包含十个高保真3D场景，总计超过18,000个任务，旨在模拟复杂的现实世界场景。DOZE的特点包括多种移动人形障碍、广泛的开放词汇对象、多样化的独特属性对象及有价值的文本提示。数据集不仅提供代理与静态障碍物之间的碰撞检查，还增加了代理与移动障碍物之间碰撞检测的功能，从而评估代理在动态环境中的避障能力。DOZE的应用领域主要集中在解决零样本物体导航问题，特别是在动态环境中的导航效率、安全性和物体识别准确性方面。

DOZE is a dataset designed for open-vocabulary zero-shot object navigation in dynamic environments, developed by Peking University. This dataset includes ten high-fidelity 3D scenes, totaling over 18,000 tasks, aimed at simulating complex real-world scenarios. Key features of DOZE encompass diverse moving humanoid obstacles, a wide range of open-vocabulary objects, varied objects with unique attributes, and valuable textual prompts. In addition to offering collision checks between the agent and static obstacles, the dataset also integrates collision detection functionality between the agent and moving obstacles, thus enabling the evaluation of the agent's obstacle avoidance capabilities in dynamic environments. The primary application scenarios of DOZE focus on resolving zero-shot object navigation problems, especially concerning navigation efficiency, safety, and object recognition accuracy in dynamic environments.