AI capsule endoscopy navigates stomach with 97% coverage autonomously

A team of researchers from Shandong University, Chinese University of Hong Kong, and National University of Singapore has published a paper on arXiv introducing AL-DRL (Anatomical Landmark-Guided Deep Reinforcement Learning) for autonomous gastric navigation using wireless capsule endoscopy. Traditional capsule endoscopy suffers from incomplete mucosal coverage and poor transferability across patients. AL-DRL solves this by leveraging a lightweight edge-contour-depth fusion module that extracts stable, low-dimensional landmark coordinates rather than processing full high-definition video streams. This design bridges the sim-to-real gap, enabling policy transfer from simulation to physical devices.

In simulation tests across eight patient-derived models, AL-DRL achieved over 97% coverage within 50 seconds, significantly outperforming standard algorithms like PPO, SAC, and DQN. To handle real-world disturbances, the team built a two-stage sim-to-real pipeline with an adaptive dynamic programming controller. In ex-vivo experiments, the system reached a mean coverage of 87% while reducing procedure time by 53% compared to expert manual control. This advancement could make wireless capsule endoscopy more reliable and clinically viable for gastric diagnostics.