Combinatorial Robot Planning Tasks

该数据集是为了评估机器人根据语言指令操作积木的能力而构建的，旨在测试组合泛化的能力。数据集分为已知语言指令（占70%）和未知语言指令（占30%），在每个环境迭代中，积木和碗的位置都是完全随机化的。该数据集包含了20万个示例视频用于训练，任务重点在于机器人操作任务中的组合泛化能力。

This dataset is constructed to evaluate robots' ability to manipulate building blocks following linguistic instructions, aiming to test compositional generalization capabilities. The dataset is split into two categories: known linguistic instructions (accounting for 70%) and unknown linguistic instructions (accounting for 30%). The positions of both building blocks and bowls are fully randomized during each environment iteration. The dataset contains 200,000 sample videos for training, and the task focuses on the compositional generalization ability in robotic manipulation tasks.