CHAIN (Causal Hierarchy of Actions and Interactions)
收藏资源简介:
CHAIN是由新加坡科技设计大学等机构联合开发的交互式3D物理推理基准测试,包含109个机械拼图和空间堆叠任务,分为三个难度等级。数据集通过Unity引擎和Python 3D环境构建,采用多视角观察和色彩提示控制方案,重点评估智能体对几何约束、接触力学和多体支撑关系的理解能力。该基准旨在推动物理基础智能体的发展,解决传统视觉语言模型在长程规划、可行性推理和动作执行方面的局限性,适用于具身智能、交互设计和自动化操作等领域的研究。
CHAIN is an interactive 3D physical reasoning benchmark co-developed by institutions including the Singapore University of Technology and Design, consisting of 109 mechanical puzzle and spatial stacking tasks divided into three difficulty levels. The dataset is constructed using the Unity engine and Python-based 3D environments, adopting a multi-view observation and color cue control scheme, with a focus on evaluating an AI agent's understanding of geometric constraints, contact mechanics, and multi-body support relationships. This benchmark aims to advance the development of physically grounded AI agents, addressing the limitations of traditional vision-language models in long-horizon planning, feasibility reasoning, and action execution, and is applicable to research in embodied intelligence, interactive design, automated manipulation, and other related fields.



