Mujoban

该数据集以“Mujoban”为基准任务，要求机器人根据2D推箱子谜题生成的3D仓库进行布局。这一任务融合了视觉、抽象和物理推理的挑战。在其他信息中，报告了不同配置和难度级别下的成功率，结果显示模块化强化学习优于单一强化学习方法。该数据集的规模包括在各个难度级别（简单、中等、困难）的512个随机关卡中，任务的目的是解决3D推箱子谜题。

This dataset takes "Mujoban" as its benchmark task, where robots are tasked with creating layouts for 3D warehouses generated from 2D Sokoban puzzles. This task combines challenges in visual perception, abstract reasoning, and physical reasoning. Additionally, success rates under various configurations and difficulty levels were reported, with results showing that modular reinforcement learning outperforms single-mode reinforcement learning methods. The dataset comprises 512 randomly generated levels across three difficulty tiers: simple, medium, and hard, with the core objective of solving the 3D Sokoban puzzles.