Planar Manipulator Dataset

该数据集由 90 000 个彩色视频组成，这些视频显示了平面机器人机械手执行关节式操作任务。更准确地说，机械手抓住一个随机颜色和大小的圆形物体，并将其放置在一个同样随机颜色和大小的方形物体/平台上。对象的初始配置（位置、大小和颜色）是在生成过程中随机采样的。与移动 MNIST 数据集等其他数据集不同，这些样本包含所描述的面向目标的任务，使其更适合测试 ML 模型的预测能力。例如，可以将其用作玩具数据集来研究深度神经网络的容量和输出行为，然后再在真实数据上进行测试。

This dataset consists of 90,000 color videos showing a planar robotic manipulator performing articulated manipulation tasks. More specifically, the manipulator grasps a circular object of random color and size, and places it onto a square object/platform of equally random color and size. The initial configurations (position, size, and color) of the objects are randomly sampled during generation. Unlike other datasets such as the Moving MNIST dataset, these samples feature the described goal-oriented tasks, making them more suitable for testing the predictive capabilities of ML models. For example, they can serve as a toy dataset to study the capacity and output behavior of deep neural networks prior to testing on real-world data.