Image Quality Assessment for Embodied AI

image quality assessment for Human->Machine->Robotic Visual Systems

Center of AI Evaluation, Shanghai AI Lab

Embodied AI, as a bridge connecting external and internal realities, has developed rapidly in recent years. Relying on its ability to interact with the physical environment, Embodied AI has been applied to simple scenarios, but it is not yet capable of handling complex environments like autonomous driving and wilderness exploration. Unlike traditional robotics driven by fixed algorithms, Embodied AI collects signals from the Real-world and is therefore susceptible to distortions. For example, a pick-and-place task may be successfully debugged in the laboratory, but it may fail in Real-world tasks due to slight lens defocusing or shaking. Therefore, the preferences of Embodied AI should be analyzed to filter out these low-quality images.

Considering this issue, we first attempt to implement Image Quality Assessment (IQA) metrics to extend the application of Embodied AI.

🤗 Datasets Download | 📚 Paper | 📈Benchmark

Release

• [2026/1/31] 🔥 Source quality annotations of EmbodiedIQA are released!
• [2026/1/26] 🔥 EmbodiedIQA is accepted by ICLR!
• [To Do] [ ] Real-world data.

Installing

Citation

If you find our work interesting, please feel free to cite our paper:

@misc{li2025imagequalityassessmentembodied,
      title={Image Quality Assessment for Embodied AI}, 
      author={Chunyi Li and Jiaohao Xiao and Jianbo Zhang and Farong Wen and Zicheng Zhang and Yuan Tian and Xiangyang Zhu and Xiaohong Liu and Zhengxue Cheng and Weisi Lin and Guangtao Zhai},
      year={2025},
      eprint={2505.16815},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.16815}, 
}