Visuomotor affordance learning (VAL) robot interaction dataset

该数据包含 Sawyer 机器人与各种对象交互的大约 2500 条轨迹（带有图像和动作）。数据集中的示例显示在相邻的视频中。我们提供两个版本的 VAL 数据集——一个带有低分辨率图像 (1.4 GB)，另一个带有高分辨率图像 (162 GB)。这两个版本的数据量和格式相同；区别仅在于图像观察质量。较小的数据集，具有 48x48x3 图像，可用于例如。离线 RL，可直接下载：https://drive.google.com/file/d/1UuWANkVtWLg4egIK2LB_YCKuF87rMQ1H/view?usp=sharing表示学习，可在此 Google 驱动器文件夹中找到：https://drive.google.com/drive/folders/1kD9kyP7-RlIrSnuN7rpEASAGWp5qnNov?usp=sharing 要下载更大的数据集，我们建议使用 https://rclone.org/数据被分类到几个文件夹中。共有300个文件和2500条轨迹。 - fixed_drawer - 人工控制的演示数据打开和关闭抽屉。 (~10%) - fixed_pnp - 人工控制的演示数据拾取对象。 (~10%) - fixed_pot - 与锅和盖子交互的人工控制演示数据。 (~10%) - fixed_tray - 人工控制的演示数据拾起物体并将其放入托盘中。 (~10%) - 一般 - 进一步收集的人工控制的演示数据具有最大的多样性和变化。 (~40%) - onpolicy_eval - RL 策略收集的评估数据。 (~10%) - onpolicy_expl - RL 策略收集的探索数据。 (~10%)

This dataset contains approximately 2,500 trajectories (paired with images and actions) of the Sawyer robot interacting with various objects. Examples from the dataset are shown in adjacent videos. We provide two versions of the VAL dataset: one with low-resolution images (1.4 GB) and another with high-resolution images (162 GB). The two versions have identical data volume and format; the only difference is the quality of image observations. The smaller dataset, featuring 48x48x3 images, can be used for tasks such as offline RL and representation learning, and is directly downloadable at: https://drive.google.com/file/d/1UuWANkVtWLg4egIK2LB_YCKuF87rMQ1H/view?usp=sharing. The larger dataset is available in this Google Drive folder: https://drive.google.com/drive/folders/1kD9kyP7-RlIrSnuN7rpEASAGWp5qnNov?usp=sharing. For downloading the larger dataset, we recommend using https://rclone.org/.

The data is categorized into several folders. In total, there are 300 files and 2,500 trajectories:
- fixed_drawer: Manually controlled demonstration data for opening and closing drawers (~10%)
- fixed_pnp: Manually controlled demonstration data for picking up objects (~10%)
- fixed_pot: Manually controlled demonstration data for interacting with pots and their lids (~10%)
- fixed_tray: Manually controlled demonstration data for picking up objects and placing them into trays (~10%)
- general: Further collected manually controlled demonstration data with the largest diversity and variation (~40%)
- onpolicy_eval: Evaluation data collected by RL policies (~10%)
- onpolicy_expl: Exploration data collected by RL policies (~10%)

搜集汇总

背景与挑战

背景概述

该数据集包含约2500条Sawyer机器人与多种物体交互的轨迹，提供低分辨率和高分辨率两个版本，数据按任务类型分类，涵盖人工控制演示和强化学习策略收集的示例。