Reproduce “EndoWAM: A Grounded World-Action Model for Generalizable Endoscopic Navigation” 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
Jinsong Lin, Zikang Pan, Wanhao Liu, Chi Kit Ng, Liangjing Shao, Zihang Yu, Ziyu Wang, Yin Wang and 5 more
EndoWAM, a World Action Model, provides generalizable autonomous navigation for flexible endoscopes by predicting future anatomical dynamics and employing a "future grounding" strategy to maintain target focus. The system achieved an 80.2% average success rate across diverse endoscopic procedures and demonstrated 90.1% zero-shot generalization to unseen conditions, while maintaining a real-time control frequency of 7.5 Hz.
Key insights
- Reconstructing future target regions from predictive denoising features is an effective strategy for robustly coupling visual dynamics, target-aware attention, and multi-step action generation.
- Viewpoint augmentation during training is crucial for enabling robust generalization to unseen orientations encountered in endoscopic navigation.
- Integrating predictive dynamics and future grounding benefits during training, while performing action generation from a single denoising pass at inference, allows for real-time control without significant latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.