Reproduce “Image Quality Assessment for Embodied AI” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chunyi Li, Jiaohao Xiao, Jianbo Zhang, Farong Wen, Zicheng Zhang, Yuan Tian, Xiangyang Zhu, Xiaohong Liu and 3 more
Researchers from Shanghai Jiao Tong University and Shanghai AI Lab introduce Image Quality Assessment (IQA) for Embodied AI, establishing a theoretical framework and a large-scale database. This work quantitatively demonstrates that existing human-centric and general machine-oriented IQA methods are insufficient for predicting robot performance in real-world tasks under visual distortions, providing a foundational resource for robust robotic perception.
Key insights
- Embodied AI agents exhibit distinct sensitivities to visual distortions compared to human visual systems, with certain distortions severely impacting robotic performance at low intensity while others have minimal effect.
- There is a low correlation (SRCC around 0.3 for VLMs, 0.25 for VLAs) between the performance of different VLM and VLA models under distortion, indicating high diversity in model responses and the necessity for multi-model annotations.
- A significant disconnection exists between a robot's cognitive understanding, its decision-making, and its final physical execution, validating that assessing only perception or cognition is insufficient for predicting embodied system success.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.