Fig. Proposed Bab_Sak Robotic Intubation System (BRIS) featuring the Actual Setup (Left) and its 3D Rendered Model (Right)

Fig. Detailed View of the Proposed End Effector for Intubation, Highlighting Its Design and Functional Specifications

We propose Bab_Sak Robotic Intubation System (BRIS), a control-centric robotic platform that integrates learning-enabled shape-aware control with real-time monocular depth perception for fiberoptic endotracheal intubation. The core idea is to decouple navigation and control stability from anatomy-aware placement verification: learning-based closed-loop control ensures predictable and intuitive distal instrument motion under tendon nonlinearities and airway contact, while perception modules provide interpretable, real-time awareness of tube depth relative to the carina. This design preserves human authority and clinical workflow compatibility while enabling objective, safety-critical feedback during intubation. The main contributions are:

• To our knowledge, BRIS is the first robotic fiberoptic intubation system to integrate independent robotic tube advancement with real-time, perception-based tube-to-carina depth verification, addressing a critical but previously uninstrumented dimension of intubation safety.

• A learning-enabled closed-loop control framework for tendon-driven fiberoptic instruments, which leverages real-time shape sensing and latent dynamics modeling to provide stable Cartesian distal tip control under frictional hysteresis, nonlinear coupling, and airway contact.

• An anatomy-aware monocular depth estimation framework that classifies airway regions and provides interpretable guidance for safe tube placement, enabling depth awareness without dedicated depth sensors, structured illumination, or patient-specific calibration.

• A compact, clinically compatible robotic hardware platform, comprising a four-way steerable fiberoptic bronchoscope, independent endotracheal tube advancement, and a camera-augmented mouthpiece designed to integrate seamlessly with standard intubation workflows.

• Comprehensive experimental validation on high-fidelity airway mannequins, demonstrating reliable navigation, controlled tube placement, and safety-preserving behavior across both standard and constrained airway scenarios.

Fig. A short history of backbone shape measurements from the ShapeSensing™ module is encoded by a temporal convolutional network into a compact latent deformation state. This representation, together with the current tip state and actuator inputs, is used by a learned dynamics model to predict future motion. During execution, the learned model is embedded within a nonlinear model predictive control framework to generate stable, closed-loop control commands for precise fiberoptic navigation. 

BRIS was evaluated on high-fidelity airway training mannequins under both standard and constrained airway conditions that replicate realistic anatomical variability, tissue compliance, and insertion resistance. The learning-enabled control system was trained using joystick-based teleoperation data that included straight insertions, sharp bends, and contact-rich airway interactions, and executed in real time using a model predictive control framework. Across 48 complete intubation trials, the system achieved reliable tracheal access and controlled tube placement, with safety-preserving behavior maintained even in constrained airways. Monocular depth estimation enabled real-time awareness of tube position relative to the carina, achieving a mean depth estimation error of approximately 2–3 mm, which was sufficient for reliable airway zone classification. Overall, 98% of trials resulted in final tube placement within a clinically accepted safety range relative to the carina, without requiring dedicated depth sensors, structured illumination, or patient-specific calibration.