NavDP
收藏资源简介:
NavDP数据集是一个大规模的导航数据集,包含超过56,000个导航轨迹和1,000万个渲染的RGB-D图像,覆盖超过360公里的导航距离。数据集是在模拟环境中生成的,使用了一种高效的数据生成流程,每天每个GPU可以生成约2,500个轨迹,比现实世界的数据收集效率高出20倍。数据集涵盖了从3D-Front和Matterport3D选择的超过1,200个场景,每个场景有100对起点和目的地。NavDP数据集的创建旨在为机器人提供导航技能的训练数据,帮助机器人在不同的场景和机器人类型中实现零样本迁移。数据集的应用领域是机器人导航,旨在解决动态开放世界环境中的导航挑战,帮助机器人实现跨不同形态和场景的零样本泛化。
The NavDP dataset is a large-scale navigation dataset containing over 56,000 navigation trajectories and 10 million rendered RGB-D images, covering a total navigation distance of more than 360 kilometers. It is generated in a simulated environment with an efficient data generation pipeline, which can produce approximately 2,500 trajectories per GPU per day, 20 times more efficient than real-world data collection. The dataset includes over 1,200 scenes selected from 3D-Front and Matterport3D, with 100 start-goal pairs per scene. The NavDP dataset is developed to provide training data for robot navigation skills, enabling robots to achieve zero-shot transfer across various scenarios and robot types. Its application focuses on robot navigation, aiming to address navigation challenges in dynamic open-world environments and help robots realize zero-shot generalization across different robot morphologies and scenes.

- 1NavDP: Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance上海人工智能实验室 · 2025年



