BRIS: Advanced Robotic Intubation Systems

• BRIS is a family of robotic intubation systems that combine endoscopic perception, anatomical guidance, and closed-loop control for automated airway management.
• The platforms integrate advanced imaging, segmentation-driven navigation, and sensor fusion to safely plan and execute tracheal and nasotracheal intubation.
• System evaluations show 100% trial success and significant reductions in insertion force compared to manual methods, underlining its clinical potential.

Searching arXiv for the cited BRIS-related papers and closely related robotic intubation work. The Robotic Intubation System (BRIS) denotes a family of robotic platforms for airway management that couple endoscopic perception, anatomical guidance, and robot control to support or automate tracheal or nasotracheal intubation. Across the cited literature, BRIS is associated with at least three tightly related technical trajectories: a Bot-assisted Robotic Intubation System emphasizing segmentation-driven surgical navigation (Wang et al., 2023), a platform for autonomous nasotracheal intubation using force-instrumented demonstrations and a Transformer-based recurrent policy (Tian et al., 3 Aug 2025), and the Bab_Sak Robotic Intubation System, a compact human-in-the-loop system for fiberoptic-guided endotracheal intubation with learning-enabled teleoperation and monocular depth-based placement guidance (Gupta et al., 26 Dec 2025). A related perception module for landmark localization using Deformable DeTR with Semantic-Aligned-Matching was also developed for robotic nasal airway intubation and integrated into the BRIS pipeline (Liu et al., 2023). Taken together, these works define BRIS as a research program in robotic airway intervention centered on safe navigation, anatomy-aware perception, and closed-loop control.

1. Terminology, scope, and procedural targets

In the cited work, BRIS is used in the context of robot-assisted tracheal intubation, autonomous nasotracheal intubation (NTI), and fiberoptic-guided endotracheal intubation. The procedural target is the establishment of an artificial airway while reducing failure modes associated with manual airway instrumentation, including difficult anatomical access, high contact forces, and inaccurate depth placement (Tian et al., 3 Aug 2025, Gupta et al., 26 Dec 2025).

The 2023 segmentation paper describes the BRIS vision module as part of a Bot-assisted Robotic Intubation System in which a segmentation network is integrated as a ROS node and used for path planning, collision avoidance, and tool guidance (Wang et al., 2023). The 2025 autonomous NTI paper describes a BRIS platform built around a KUKA iiwa 7-DOF lightweight manipulator, an endoscopic view, and a prosthesis embedded with force sensors, with autonomy restricted to clinically relevant insertion DoFs (Tian et al., 3 Aug 2025). The later 2025 paper presents the Bab_Sak Robotic Intubation System (BRIS) as a compact human-in-the-loop platform integrating a four-way steerable fiberoptic bronchoscope, an independent endotracheal-tube advancement mechanism, and a camera-augmented mouthpiece compatible with standard clinical workflows (Gupta et al., 26 Dec 2025).

These uses are not identical. A plausible implication is that “BRIS” functions less as a single frozen hardware configuration than as a label for a sequence of related robotic intubation systems that share a common emphasis on anatomy-aware visual perception and closed-loop assistance or autonomy.

2. Hardware architectures and actuation strategies

The BRIS-related systems differ substantially in hardware embodiment. In the autonomous NTI configuration, the robot arm is a KUKA iiwa 7-DOF lightweight manipulator, with each joint instrumented with position and joint-torque sensing; the end-effector flange carries a 6-DoF pose estimate and an integrated 3D force/torque sensor (Tian et al., 3 Aug 2025). Vision consists of Camera 1, a clinical endoscope providing the “eye-in-hand” view of the extraluminal portion of the tube at 30 fps, and Camera 2, an external viewpoint used for recording only (Tian et al., 3 Aug 2025). Motion is restricted during both tele-operation and autonomy to three DoFs: planar $(x,z)$ translation and rotation about the $y$-axis (tube pitch), matching the clinically relevant insertion DoFs (Tian et al., 3 Aug 2025).

The same NTI system incorporates a multi-material 3D-printed nasal/oropharyngeal model with embedded force sensing: one 3D sensor at the nostril entrance $(F_x,F_y,F_z)$ plus two 1D sensors in the sphenoid $(F_1)$ and pharyngeal $(F_2)$ regions (Tian et al., 3 Aug 2025). End-effector force/torque is read at 500 Hz, prosthesis force sensors are logged at 200 Hz, and segmentation outputs are delivered to the RACCT model at 5 Hz (Tian et al., 3 Aug 2025).

The Bab_Sak BRIS adopts a distinct architecture centered on a four-way steerable fiber-optic bronchoscope (FOB) (Gupta et al., 26 Dec 2025). The bronchoscope has total length 1500 mm, outer diameter 6 mm $(\pm 0.1\ \mathrm{mm})$, and inner lumen diameter 4 mm $(\pm 0.1\ \mathrm{mm})$, with coaxial compatibility with standard adult ET tubes (7.5–8.5 mm) (Gupta et al., 26 Dec 2025). Sections 1 and 3–4 use Pebax 7233, while the articulation zone uses Pebax 3533; steering is implemented with 4 × 0.3 mm stainless steel tendons in a differential-tendon configuration driven by 4 × Dynamixel XM430-W210-T servos, enabling continuous orientation control of the distal tip over the unit sphere $S^2$ (Gupta et al., 26 Dec 2025). Shape sensing is provided by a Fiber-Optic Shape Sensing (FOSS) module reconstructing $N=48$ backbone points in real time (Gupta et al., 26 Dec 2025).

The Bab_Sak end-effector includes a dual lead-screw architecture with separate platforms for FOB insertion/retraction and ET-tube linear translation, each driven by NEMA 17 stepper motors with microstepping for sub-millimeter resolution (Gupta et al., 26 Dec 2025). A camera-augmented mouthpiece based on Macintosh blade curvature integrates a 2.8 mm CMOS camera with 120° FOV (Gupta et al., 26 Dec 2025). The hardware is mounted to a UR10-series robotic arm, and the software stack includes Jetson AGX and ROS Master (Linux) linked to robotic arm, Dynamixel servos, and stepper drivers (Gupta et al., 26 Dec 2025).

The 2023 BRIS vision work focuses less on manipulators and more on the imaging subsystem. Its vision module assumes a 1080p endoscopic sensor at 30 fps with latency <10 ms, and inference on NVIDIA Jetson AGX Xavier (GPU) with average segmentation time ~30 ms/frame (Wang et al., 2023). This yields an end-to-end vision-to-actuation latency target of $\leq 50\ \mathrm{ms}$ (Wang et al., 2023).

3. Perception modules: segmentation, landmarks, and depth

A defining feature of BRIS research is the use of specialized perception modules tuned to airway anatomy.

For organ-level understanding, the 2023 work develops domain adaptive Sim-to-Real segmentation of oropharyngeal organs—specifically the uvula, epiglottis, and glottis—for robot-assisted intubation (Wang et al., 2023). To address limited real endoscopic data, it constructs a photo-realistic virtual oropharyngeal phantom in the Simulation Open Framework Architecture (SOFA) framework. Organ meshes are generated in Blender from CT-derived surfaces, imported into SOFA as TetrahedronFEMObject, and modeled using linear tetrahedral elements (P1 FEM) (Wang et al., 2023). The material model is Neo-Hookean hyperelasticity with Young’s modulus $y$0 for uvula, epiglottis, and glottis, Poisson ratio $y$1, and Rayleigh damping $y$2 (Wang et al., 2023). Virtual endoscopy uses focal length 18 mm, FOV $y$3, and resolution $y$4 px, with Phong shading and four point lights (Wang et al., 2023).

The segmentation network is DeepLab v3+ with ResNet-101 encoder, output stride = 16, ASPP rates $y$5, and a 4-channel softmax for background plus three organs (Wang et al., 2023). Domain adaptation combines an IoU-Ranking Blend (IRB) strategy with CycleGAN-style style transfer using generators $y$6 and discriminators $y$7 (Wang et al., 2023). The per-class IoU is defined as

$y$8

and the IRB allocation rule is

$y$9

with $(F_x,F_y,F_z)$0 for $(F_x,F_y,F_z)$1 (Wang et al., 2023). A ranking loss

$(F_x,F_y,F_z)$2

is used to encourage the network to close the performance gap between classes (Wang et al., 2023).

For landmark-level detection, the 2023 nasal intubation work proposes a Deformable Detection Transformer augmented with Semantic-Aligned-Matching (SAM) to detect nostrils and glottis (Liu et al., 2023). The backbone is ResNet-50, feature maps are extracted at strides $(F_x,F_y,F_z)$3, the decoder uses $(F_x,F_y,F_z)$4 queries of dimension $(F_x,F_y,F_z)$5, and cross-attention is implemented with multi-scale deformable attention using $(F_x,F_y,F_z)$6 sampling points per level (Liu et al., 2023). The SAM module predicts a reference box

$(F_x,F_y,F_z)$7

and fuses original and salient ROI-derived query embeddings using a learnable gate

$(F_x,F_y,F_z)$8

yielding

$(F_x,F_y,F_z)$9

The loss follows standard DETR-style Hungarian matching with classification and box terms (Liu et al., 2023).

For depth-aware airway placement, the Bab_Sak BRIS integrates zero-shot monocular depth using the “Depth Anything” foundation model (Yang et al. 2024) to produce a dense relative-scale depth map $(F_1)$0 from live 30 fps endoscopic video (Gupta et al., 26 Dec 2025). Estimated tip-to-carina distance $(F_1)$1 is converted into Zone I, Zone II, and Zone III by a threshold-based classifier: $(F_1)$2 with example thresholds $(F_1)$3 and $(F_1)$4 (Gupta et al., 26 Dec 2025). Guidance is further supported by a passive visual servoing overlay derived from the misalignment vector

$(F_1)$5

where $(F_1)$6 is the centroid of maximal depth pixels and $(F_1)$7 is the optical center (Gupta et al., 26 Dec 2025).

In the autonomous NTI paper, the perception branch is specialized to the tube rather than directly to anatomical organs. The input branch uses endoscope image $(F_1)$8 DLO-segmentation $(F_1)$9 morphological post-processing $(F_2)$0 skeleton & curvature features, combined with 6-DoF end-effector pose and 3D robot-flange force (Tian et al., 3 Aug 2025). This suggests that BRIS perception has evolved toward task-specific representations: organ masks for navigation, landmarks for staged insertion, and tube skeleton/curvature for contact-aware policy learning.

4. Learning-enabled control and autonomy

The BRIS literature spans several control paradigms, from segmentation-assisted motion planning to learning-enabled teleoperation and full imitation-learned autonomy.

In the 2023 segmentation-integrated BRIS pipeline, segmentation masks are converted to 3D point clouds via hand-eye calibration and depth estimation (Wang et al., 2023). The 3D centroids of glottis pixels define a spatial goal for tip articulation; voxelized organ volumes from segmentation plus stereo depth feed into MoveIt! for motion planning under anatomical constraints; and uvula and epiglottis masks inform active bending of the EndoBot stylet via a closed-loop PID controlling a curvature motor (Wang et al., 2023). The software interface is organized as ROS topic /camera/image_raw → segmentation node → /segmentation/mask (Wang et al., 2023).

The autonomous NTI paper introduces Recurrent Action-Confidence Chunking with Transformer (RACCT) as the principal policy model (Tian et al., 3 Aug 2025). Inputs are embedded as

$(F_2)$1

then processed by a standard Transformer encoder using self-attention

$(F_2)$2

The decoder jointly regresses a chunk of future actions

$(F_2)$3

and associated confidences

$(F_2)$4

which are combined into an executed command

$(F_2)$5

The loss is

$(F_2)$6

with $(F_2)$7 and $(F_2)$8 (Tian et al., 3 Aug 2025). The model is expressly intended to handle complex tube-tissue interactions and partial visual observations (Tian et al., 3 Aug 2025).

The Bab_Sak BRIS uses a different learning-enabled controller: a learned forward dynamics model embedded in MPC for stable teleoperation under tendon nonlinearities and airway contact (Gupta et al., 26 Dec 2025). State, shape, and input are defined as

$(F_2)$9

A Temporal Convolutional Network (TCN) encoder maps a short window $(\pm 0.1\ \mathrm{mm})$0 with $(\pm 0.1\ \mathrm{mm})$1 (approximately 120 ms) (Gupta et al., 26 Dec 2025). A residual neural network predicts the next tip state: $(\pm 0.1\ \mathrm{mm})$2 Training uses the loss

$(\pm 0.1\ \mathrm{mm})$3

optimized with Adam, lr $(\pm 0.1\ \mathrm{mm})$4, batch 256, over 120 epochs on approximately $(\pm 0.1\ \mathrm{mm})$5 state-transition samples (Gupta et al., 26 Dec 2025).

The MPC solves a finite-horizon OCP at each time step with horizon $(\pm 0.1\ \mathrm{mm})$6, update rate = 50 Hz, and solve time $(\pm 0.1\ \mathrm{mm})$7 ms/step (Gupta et al., 26 Dec 2025). The joystick mapping translates axes to Cartesian velocity and buttons to ET tube retract, go-to-waypoint, and emergency-stop (Gupta et al., 26 Dec 2025). This is a markedly different control philosophy from RACCT: MPC is used for stable and intuitive teleoperation, whereas RACCT is used for autonomous insertion from demonstrations.

5. Safety mechanisms and quantitative performance

Safety is explicit in all BRIS variants, but implemented through different mechanisms.

In the vision-guided BRIS pipeline, end-to-end vision-to-actuation latency $(\pm 0.1\ \mathrm{mm})$8 is required to ensure real-time response (Wang et al., 2023). A watchdog monitors segmentation confidence; if mean softmax $(\pm 0.1\ \mathrm{mm})$9, the robot halts and requests clinician override (Wang et al., 2023). All control commands pass through a 2-layer safety filter checking joint limits and maximum curvature rates (Wang et al., 2023).

In the autonomous NTI system, safety begins at the data-collection stage. Demonstrations are filtered using three criteria: intubation time $(\pm 0.1\ \mathrm{mm})$0, peak prosthesis force $(\pm 0.1\ \mathrm{mm})$1, and impulse

$(\pm 0.1\ \mathrm{mm})$2

with $(\pm 0.1\ \mathrm{mm})$3, i.e. $(\pm 0.1\ \mathrm{mm})$4 (Tian et al., 3 Aug 2025). Only episodes in which all three metrics fall below 70% of these thresholds are retained for imitation learning (Tian et al., 3 Aug 2025). Real-time control uses only the robot’s flange 3D force/torque; the prosthesis sensors are used a posteriori for filtering, annotation, and evaluation (Tian et al., 3 Aug 2025).

The Bab_Sak BRIS incorporates safety into the user interface and guidance logic. Features include a hardware E-stop accessible to Operator B, software velocity limits, and depth-triggered automatic withdrawal if Zone III is entered accidentally (Gupta et al., 26 Dec 2025). The success criterion is explicit: the ET tube tip stably in Zone II (mid-trachea) (Gupta et al., 26 Dec 2025).

The principal quantitative outcomes reported across the cited BRIS works are summarized below.

System/paper	Reported task	Key outcomes
Vision BRIS (Wang et al., 2023)	Real-phantom segmentation	Overall Dice: 0.64 baseline vs 0.73 domain-adaptive; overall Pixel Acc: 0.88 vs 0.92
Autonomous NTI BRIS (Tian et al., 3 Aug 2025)	Autonomous nasotracheal intubation	100% success rate; time 9.13 s $(\pm 0.1\ \mathrm{mm})$5; peak $(\pm 0.1\ \mathrm{mm})$6 1.56 N $(\pm 0.1\ \mathrm{mm})$7 vs doctor 2.75 N $(\pm 0.1\ \mathrm{mm})$8
Bab_Sak BRIS (Gupta et al., 26 Dec 2025)	Fiberoptic-guided intubation on mannequins	48 trials; 100 % success in standard and constrained scenarios; depth MAE $(\pm 0.1\ \mathrm{mm})$9 overall

The segmentation study reports per-class gains on the real phantom: uvula Dice 0.65 to 0.72, epiglottis 0.58 to 0.68, glottis 0.70 to 0.78, and overall 0.64 to 0.73 under the domain-adaptive pipeline (Wang et al., 2023). The conclusion states $S^2$0 (Wang et al., 2023).

The autonomous NTI study reports that RACCT outperforms the ACT model in all aspects and achieves a 66% reduction in average peak insertion force compared to manual operations while maintaining equivalent success rates (Tian et al., 3 Aug 2025). More specifically, mean outcomes over successful trials include RACCT 100% success, time 9.13 s $S^2$1, doctor 5.03 s $S^2$2, and peak force in $S^2$3 of 1.81 N for RACCT versus 2.66 N for doctor (Tian et al., 3 Aug 2025). The paper additionally notes that impulse reductions show 35–60% lower continuous loading in RACCT vs. human and that the reductions across $S^2$4 trials per condition exceed inter-trial variability by $S^2$5, though no formal p-values were reported (Tian et al., 3 Aug 2025).

The Bab_Sak BRIS reports 48 total mannequin trials, split into 24 standard and 24 constrained scenarios (Gupta et al., 26 Dec 2025). It achieved 100 % success rate in both scenarios, with no endobronchial or esophageal misplacements (Gupta et al., 26 Dec 2025). Depth-estimation MAE was $S^2$6 overall, $S^2$7 standard, and $S^2$8 constrained, and 98 % of trials ended within $S^2$9 of target (Gupta et al., 26 Dec 2025). Ablations showed 46 % reduction in joystick command variance during distal navigation with learned MPC versus naïve teleoperation without shape sensing; turning visual guidance off in a 16-trial subset increased distal wall contacts by 52 % and corrective withdrawals by 35 % (Gupta et al., 26 Dec 2025). Tracking RMSE was < 1.5 mm in free-space trajectories and < 3 mm under soft contact (Gupta et al., 26 Dec 2025).

6. Datasets, training regimes, and experimental methodology

The BRIS literature relies on heterogeneous data sources, reflecting the scarcity of real labeled airway datasets and the safety constraints of clinical experimentation.

For segmentation, the 2023 paper explicitly motivates virtual data generation because real datasets of oropharyngeal organs are limited due to patient privacy issues (Wang et al., 2023). Its synthetic source domain is generated in SOFA from medical-grade CAD scans and CT-derived surfaces (Wang et al., 2023). The segmentation model is trained with SGD, momentum = 0.9, weight decay = $N=48$0, initial learning rate 0.01, “poly” policy $N=48$1, batch size 8 (4 synthetic + 4 blended), and 100 epochs (Wang et al., 2023). Data augmentation includes random horizontal/vertical flips, scaling [0.8–1.2], and color jitter $N=48$2 (Wang et al., 2023).

For landmark detection, two datasets are used (Liu et al., 2023). The nostril dataset is derived from the BioID face keypoints dataset with 1 521 images, resized to 640×480, split into 1 021 train / 185 val / 315 test, and automatically annotated by expanding the nostril keypoint to a $N=48$3 px box (Liu et al., 2023). The glottis dataset comes from the BAGLS segmentation dataset, with 881 nasal endoscopy frames, split into 377 train / 88 val / 416 test after converting masks to axis-aligned bounding boxes (Liu et al., 2023). All models are trained for 24 epochs with Adam, $N=48$4, $N=48$5, batch 8, in MMDetection (Liu et al., 2023).

The landmark detector achieves the following benchmarked performance. On glottis detection, Ours (SAM Deformable) reports mAP [0.5:0.95] = 0.282, [email protected] = 0.661, [email protected] = 0.270, [email protected] = 0.68, and [email protected] = 0.59 (Liu et al., 2023). On nostril detection, it reports mAP [0.5:0.95] = 0.325, [email protected] = 0.865, [email protected] = 0.142, [email protected] = 0.88, and [email protected] = 0.79 (Liu et al., 2023). GPU inference runs at ~12 fps (80 ms/image) on an NVIDIA RTX 2080Ti, with a projected path to ~40 ms/image via reduced image resolution, TensorRT fusion, and fewer queries/layers (Liu et al., 2023).

For autonomous NTI, the dataset consists of 50 human-teleoperation episodes from two experienced operators on the prosthesis, filtered using the safety criteria above (Tian et al., 3 Aug 2025). Per-timestep recordings include image, segmented skeleton and curvature, 6-DoF pose, 3-D flange force, and prosthesis forces for annotation (Tian et al., 3 Aug 2025). Images are cropped/resized to 256×256 and normalized, with a random 80/10/10 train/val/test split (Tian et al., 3 Aug 2025). Training uses Adam, learning rate = $N=48$6, batch size = 8, chunk length $N=48$7, and 20 k gradient steps, requiring approximately 1 hour on RTX A6000 (Tian et al., 3 Aug 2025).

For the Bab_Sak dynamics model, training data comprise approximately $N=48$8 state-transition samples from teleoperated maneuvers (Gupta et al., 26 Dec 2025). The optimizer is Adam, with lr = $N=48$9, batch 256, and 120 epochs (Gupta et al., 26 Dec 2025). The training and validation loss trajectories are described as exhibiting smooth decline to plateau (Gupta et al., 26 Dec 2025).

7. Limitations, misconceptions, and research directions

A common misconception would be to treat BRIS as a single monolithic system with one settled architecture. The cited literature does not support that simplification. Instead, BRIS appears as a set of related robotic intubation systems spanning segmentation-guided robotic navigation (Wang et al., 2023), autonomous nasotracheal insertion (Tian et al., 3 Aug 2025), and human-in-the-loop fiberoptic-guided endotracheal intubation with objective depth awareness (Gupta et al., 26 Dec 2025). This suggests an evolving platform concept rather than a single immutable device.

Another potential misconception is that robotic intubation research is exclusively a navigation problem. The later BRIS papers explicitly extend beyond navigation to contact-force reduction (Tian et al., 3 Aug 2025) and to objective verification of tube depth relative to the carina (Gupta et al., 26 Dec 2025). The 2025 Bab_Sak paper further states that existing robotic and teleoperated systems primarily focus on airway navigation and do not provide integrated control of ET-tube advancement or objective verification of tube depth relative to the carina (Gupta et al., 26 Dec 2025).

Observed failure modes are also explicitly documented. In the segmentation-driven BRIS pipeline, low-lighting frames cause false negatives on the transparent epiglottis, anatomical occlusions such as tongue protrusion can degrade Dice by up to 15%, and excessive blood or secretions introduce specular highlights that confuse the network (Wang et al., 2023). Proposed enhancements include a photometric pre-processing module with specular removal and adaptive histogram equalization, multi-view 3D reconstruction, tactile/force feedback integration, and online continual learning driven by uncertainty-based annotation in the OR (Wang et al., 2023).

The autonomous NTI system is limited by validation on a rigid phantom; the paper states that animal and cadaver studies are needed to assess tissue compliance variations (Tian et al., 3 Aug 2025). It also notes that no real-time prosthesis-sensor feedback is used in closed loop and suggests future incorporation of active force-control for on-the-fly safety stopping (Tian et al., 3 Aug 2025). Generalization to patient-specific anatomies and other tubular insertion tasks is also identified as requiring domain-adaptive vision modules and expanded demo datasets (Tian et al., 3 Aug 2025).

The Bab_Sak BRIS, while demonstrating reliable navigation and controlled tube placement on high-fidelity airway mannequins, remains a mannequin-validated platform (Gupta et al., 26 Dec 2025). Its core contribution is to make fiberoptic-guided intubation more compatible with standard clinical workflow while adding real-time anatomy-aware guidance and depth awareness, but the cited text does not report human or cadaver deployment (Gupta et al., 26 Dec 2025).

Across these works, the central research direction is consistent: combine geometry-aware or anatomy-aware perception, explicit safety criteria, and learning-enabled control to reduce operator burden and tissue loading while improving consistency of airway access and final tube placement.