Depth-enabled fiber-bundle endoscopy for in situ intraluminal robotic navigation

Three-dimensional perception is essential for intraluminal robotic navigation, where depth estimation provides geometric feedback for safe advancement. Fiber-bundle endoscopy provides sub-millimeter, flexible imaging for miniature probes but lacks reliable depth recovery due to artifacts and missing ground truth. We present a weakly supervised framework that augments fiber-bundle imaging with depth by combining domain translation, pseudo-depth supervision, and structure-aware constraints. Trained without paired labels or external sensors, the network infers depth directly from raw images and reconstructs lumen topology for inclination-angle guidance. Integrated with magnetic actuation, the method enables efficient closed-loop navigation in curved luminal phantoms, demonstrating depth-enabled fiber-bundle endoscopy for in situ robotic control.