fly1113/UniWM_Dataset

UniWM数据集是一个用于视觉导航任务的数据集，托管在Hugging Face上（fly1113/UniWM_Dataset）。该数据集包含多个子数据集，用于训练和评估统一世界模型（UniWM）。具体子数据集包括：go_stanford、recon、sacson和scand，这些数据集用于训练和评估；tartandrive仅用于未见过的评估；1XHumanoid用于人形机器人导航。每个子数据集以轨迹文件夹形式组织，每个轨迹文件夹包含一系列自我中心视角的帧图像（如0.jpg, 1.jpg等）和一个traj_data.pkl文件，该文件存储了轨迹的每步元数据（如动作、姿态）。数据集支持UniWM模型的研究，该模型整合了记忆增强的规划和视觉预见，以提高导航的稳定性和连贯性。The UniWM Dataset is a dataset for visual navigation tasks, hosted on Hugging Face (fly1113/UniWM_Dataset). This dataset comprises multiple sub-datasets for training and evaluating the Unified World Model (UniWM). Specifically, the sub-datasets include go_stanford, recon, sacson, and scand, which are used for both training and evaluation; tartandrive is only for unseen evaluation; and 1XHumanoid is for humanoid robot navigation. Each sub-dataset is organized in the form of trajectory folders. Each trajectory folder contains a series of egocentric view frame images (e.g., 0.jpg, 1.jpg, etc.) and a traj_data.pkl file, which stores step-by-step metadata of the trajectory such as actions and poses. This dataset supports research on the UniWM model, which integrates memory-augmented planning and visual foresight to improve the stability and coherence of navigation.