PAI-Bench
收藏资源简介:
PAI-Bench是由佐治亚理工学院和卡内基梅隆大学的研究团队构建的综合性基准数据集,旨在系统评估人工智能模型在物理AI领域的感知与预测能力。该数据集包含2,808个高质量的真实世界视频案例,覆盖自动驾驶、机器人、工业应用等多个子领域,数据主要来源于行车记录仪等现实捕捉设备。其构建过程采用了先进的多模态大语言模型进行初始标注与问答对生成,并辅以严格的人工校验与修正。该数据集的核心应用在于为视频生成、条件视频生成及视频理解任务提供统一的评估框架,旨在解决当前模型在物理合理性、因果推理及复杂动态建模方面的关键不足,推动面向真实物理世界交互的AI系统发展。
PAI-Bench is a comprehensive benchmark dataset constructed by research teams from the Georgia Institute of Technology and Carnegie Mellon University, designed to systematically assess the perception and prediction abilities of artificial intelligence models in the domain of Physical AI. This dataset comprises 2,808 high-quality real-world video cases spanning multiple subfields including autonomous driving, robotics, and industrial applications, with the majority of data sourced from real-world capture devices such as dashcams. During its construction, advanced multimodal large language models were employed for initial annotation and question-answer pair generation, paired with rigorous manual verification and correction. The core application of this dataset is to provide a unified evaluation framework for video generation, conditional video generation, and video understanding tasks, with the goal of addressing critical limitations of existing models in physical plausibility, causal reasoning, and complex dynamic modeling, thereby advancing the development of AI systems oriented toward real physical world interactions.
Physical AI Bench (PAI-Bench) 数据集概述
数据集简介
Physical AI Bench (PAI-Bench) 是一个用于评估物理AI生成与理解能力的综合性基准测试套件。该基准覆盖了包括自动驾驶、机器人技术、工业(智能空间)以及以自我为中心的日常场景在内的多种物理场景。
核心子任务
PAI-Bench 包含三个子任务:
- PAI-Bench-G (视频生成):评估世界基础模型在给定当前状态和控制信号的情况下预测未来状态的能力。
- PAI-Bench-C (条件视频生成):专注于世界模型在更复杂控制信号(如边缘、分割掩码、深度等)下的生成能力。
- PAI-Bench-U (视频理解):评估对物理场景的理解能力。
数据集详情
各子任务对应的数据集如下:
| 任务 | 数据 | 使用说明 |
|---|---|---|
| PAI-Bench-G | 🤗 physical-ai-bench-generation | Link |
| PAI-Bench-C | 🤗 physical-ai-bench-conditional-generation | Link |
| PAI-Bench-U | 🤗 physical-ai-bench-understanding | Link |
排行榜
排行榜可通过以下地址访问:🤗 physical-ai-bench-leaderboard。
技术信息
- Python版本:3.10
- 许可证:MIT
- 相关论文:arXiv:2512.01989
- 关联机构:Georgia Tech, CMU
引用
若在研究中使用了 Physical AI Bench,请引用: bibtex @misc{zhou2025paibenchcomprehensivebenchmarkphysical, title={PAI-Bench: A Comprehensive Benchmark For Physical AI}, author={Fengzhe Zhou and Jiannan Huang and Jialuo Li and Deva Ramanan and Humphrey Shi}, year={2025}, eprint={2512.01989}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2512.01989}, }




