Reproduce “Embodied Image Compression” 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, Rui Qing, Jianbo Zhang, Yuan Tian, Xiangyang Zhu, Zicheng Zhang, Xiaohong Liu, Weisi Lin and 1 more
Researchers from Shanghai Jiao Tong University, Shanghai AI Lab, and Nanyang Technological University introduce and formally define "Embodied Image Compression," a new problem distinct from human and generic machine vision compression. They establish EmbodiedComp, a closed-loop benchmark for evaluating image compression in robotic manipulation tasks under ultra-low bandwidth. Their findings demonstrate that existing codecs fail under these conditions, with agents exhibiting a unique "robust-then-cliff" performance degradation, underscoring the need for new compression strategies for practical Embodied AI deployment.
Key insights
- RVS-oriented metrics (task success rate, steps) show very weak correlation with compression bitrate (below 0.20), indicating traditional HVS/MVS quality metrics are poor predictors of Embodied AI performance at ultra-low bitrates.
- Embodied agents exhibit a "robust-then-cliff" degradation pattern, maintaining performance at light compression but collapsing abruptly below a critical bitrate threshold (around 0.04 bpp).
- Codecs performing well for HVS and MVS often underperform simpler or traditional codecs when evaluated on RVS metrics, suggesting existing advanced compressors may over-fit to HVS/MVS priors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.