💻 computer science

Layered Risk Mapping for Autonomous Patient Transport in Expeditionary Medical Facilities

This paper presents a layered risk mapping framework that fuses heterogeneous environmental hazards using a Noisy-OR model to enable safe, autonomous patient transport in expeditionary medical facilities, significantly reducing collision rates and improving obstacle clearance compared to risk-unaware baselines.

Original authors: Lorena Maria Genua, Sarvesh Prajapati, Damla Leblebicioglu, Taşkın Padır

Published 2026-07-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lorena Maria Genua, Sarvesh Prajapati, Damla Leblebicioglu, Taşkın Padır

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where robots aren't just vacuuming your living room or delivering pizza, but are actually helping to save lives in the most chaotic, messy, and dangerous places on Earth. This paper lives in the exciting corner of robotics known as "autonomous navigation," where machines learn to move on their own without bumping into things or people. To understand this, you need to know two simple things: first, robots usually get confused when the ground changes from smooth tile to muddy grass, or when a tent flap flutters in the wind; second, in places like outbreak field hospitals, human helpers are stretched so thin that every time they carry a patient, they risk getting sick themselves and using up precious protective gear. The big question here is: Can we build a robot wheelchair that is brave enough to drive through mud, dodge moving people, and climb small hills all by itself, so human doctors can stay safe and keep treating patients?

The researchers behind this study decided to build a "smart brain" for a wheelchair that acts like a super-attentive tour guide. Instead of just looking for walls to avoid, this guide creates a special "risk map" that combines four different kinds of dangers into one big picture. Think of it like a video game where the screen doesn't just show obstacles, but also glows red for steep slopes, blue for slippery mud, and flashes for moving people. The team calls this a "layered risk mapping framework." They tested this system in a computer simulation that looked like a temporary hospital made of tents, gravel, and mud, and then they actually drove a real, commercial electric wheelchair through indoor and outdoor courses to see if it worked.

The main discovery is that this new "layered" way of thinking about danger works much better than the old ways. In their computer tests, when the robot didn't know about the risks (the "no risk" mode), it crashed into things or people more than 73% of the time. But when they turned on their new risk-mapping brain, the crash rate dropped to under 32%. The robot also learned to keep a safe distance of about 0.50 meters from obstacles, which is a huge improvement over the risky 0.14 to 0.17 meters it kept when it was ignoring the dangers.

The secret sauce they used is a math trick called "Noisy-OR." Imagine you are trying to decide if it's safe to go outside. You have four friends giving you advice: one says "It's too steep," another says "The ground is too muddy," a third says "There's a dog," and a fourth says "There's a wall." If you just took the loudest voice (the "Maximum" method), you might miss a small but dangerous warning. If you added up all the voices, you might get confused by too much noise. The "Noisy-OR" method is like a smart judge who says, "If any of these friends says there is a problem, we treat it as a problem, but we weigh how likely each friend is to be right." This approach helped the robot stay safer than any other method they tried, keeping the highest risk it ever faced lower than the other methods, even when the environment got very crowded and messy.

The team also proved this wasn't just a computer fantasy. They strapped their software onto a real Quickie Q500M powered wheelchair and drove it through three different missions: picking up a patient from the outside, moving them between tents, and taking them to the exit. Whether the ground was a smooth indoor floor with a fake "dangerous" patch or a bumpy, grassy outdoor field, the wheelchair successfully navigated the course, swerving around hazards and sticking to its path. While the system isn't perfect yet—it sometimes got a little lost in areas with few features or when the lighting changed suddenly—the results show that this layered risk map is a solid step toward making autonomous patient transport a reality in places where human helpers are needed most.

Technical Summary: Layered Risk Mapping for Autonomous Patient Transport in Expeditionary Medical Facilities

Problem Statement
Expeditionary medical facilities, such as those deployed for outbreak responses (e.g., Ebola), operate in a unique and challenging regime that existing robotics research has not fully addressed. These environments feature semi-structured, hybrid terrains (gravel, mud, temporary boardwalks) and dynamic, PPE-constrained corridors. Routine patient transport in these settings imposes a severe operational burden: it consumes personal protective equipment (PPE), diverts clinical staff from bedside care, and increases infection risks for healthcare workers. While autonomous wheelchairs offer a potential solution to absorb this load, standard navigation systems for indoor service robots or outdoor autonomous ground vehicles (AGVs) are insufficient. Indoor systems typically rely on pre-mapped, structured environments, while outdoor systems often fail to account for the specific safety constraints of patient transport or the heterogeneity of unstructured medical facility terrain. The core challenge is achieving safe, fully autonomous patient transport across these hybrid indoor-outdoor environments without human escorts, while navigating complex hazards like terrain slope, static/dynamic obstacles, and semantic traversability.

Methodology
The authors propose a Layered Risk Mapping Framework designed to fuse heterogeneous environmental hazards into a unified probabilistic cost surface. The system operates on a commercial Quickie Q500M mid-wheel-drive powered wheelchair equipped with RGB-D cameras and an onboard computing unit.

Risk Layer Definition: The framework constructs four distinct 2D risk grids, each representing a specific hazard source:
- Terrain Slope ( $R_{slope}$ ): Derived from global elevation data (FABDEM/Copernicus GLO-30). Risk is tiered based on wheelchair accessibility regulations (ADA) and the manufacturer's operational limits (Sunrise Medical). Slopes <5° are safe; 5°–8° incur moderate risk; 8°–20° incur high risk; >20° are assigned maximum risk (imminent rollover).
- Semantic Traversability ( $R_{terr}$ ): Estimated using a fine-tuned SegFormer-B2 model on RGB-D streams. Semantic labels (e.g., grass, mud, road, water) are mapped to traversability costs based on mechanical interaction and slip risk.
- Static Obstacles ( $R_{static}$ ): Generated via RTAB-Map using fused RGB-D and inertial data to create a globally consistent 3D point cloud, discretized into an occupancy grid.
- Dynamic Obstacles ( $R_{dyn}$ ): Populated from real-time updates in the RTAB-Map stream, treating moving hazards (e.g., personnel) as lethal costs.
Probabilistic Fusion (Noisy-OR): Instead of using heuristic maximum-value blending or additive costmaps (which can dilute risks or create artificial blockades), the authors employ a Probabilistic Noisy-OR Fusion model. Each risk layer is normalized to a probability index ( $\mu_i$ ) representing the independent probability of a safety violation. The fused risk ( $\mu_{fused}$ ) is calculated as the joint probability that at least one violation occurs:
$\mu_{fused} = 1 - \prod_{i} (1 - \mu_i)$
This approach ensures that independent failure modes are compounded probabilistically rather than dominated by a single layer.
Planning Stack: The unified cost surface drives a Model Predictive Path Integral (MPPI) local planner. The planner rolls out stochastic trajectory samples under unicycle dynamics to optimize a cost function that balances progress toward the goal, directional alignment, and the fused risk surface.

Key Contributions

Novel Application Domain: The paper formalizes autonomous patient transport in expeditionary medical facilities as a distinct operational regime requiring hybrid indoor-outdoor navigation.
Layered Risk Framework: It introduces a framework that probabilistically fuses four independent hazard layers (slope, static/dynamic obstacles, semantic traversability) using a Noisy-OR model, addressing the limitations of deterministic layer blending.
Empirical Validation: The framework is validated through a paired Monte-Carlo simulation study and real-world experiments on a commercial powered wheelchair across three mission profiles: intake (triage to module), intra-facility transfer, and egress (module to discharge).

Results

Simulation: In a Gazebo simulation of an outbreak field hospital with varying hazard densities, the Noisy-OR fusion method significantly outperformed a risk-unaware baseline and other fusion strategies (Log-odds, Maximum).
- Collision Reduction: Risk-informed planning reduced collision rates from 73.3%–82.2% (No Risk baseline) to under 32% across all densities.
- Obstacle Clearance: The Noisy-OR method maintained the highest and most consistent obstacle clearance (approx. 0.50 m) compared to the baseline (0.14–0.17 m).
- Risk Exposure: High-risk zone entry rates dropped from over 44% to below 13.4%. Notably, Noisy-OR achieved the lowest conditional peak risk (0.47–0.55) across all densities, indicating that when it did traverse marginal-risk regions, the severity of exposure was tightly bounded.
Real-World Validation: The system successfully executed all three mission profiles in both controlled indoor environments and unstructured outdoor grass terrains with elevation changes. The planner dynamically routed around hazardous floor patches and terrain variations.
- Limitations: The authors note that current onboard computation (laptop-based) creates resource competition between semantic segmentation and planning. Additionally, visual SLAM (RTAB-Map) experienced tracking drift in feature-sparse outdoor environments, affecting local planning accuracy.

Significance and Claims
The paper claims that this architecture successfully meets the planning requirements of a previously unaddressed safety-critical domain. By demonstrating that a Noisy-OR fusion model can effectively manage the compounding risks of hybrid terrains and dynamic obstacles, the work validates the feasibility of autonomous patient transport in expeditionary settings. The authors emphasize that their approach prevents the "dual drain" on personnel and infection-control resources during surge conditions. The study concludes that while current limitations exist regarding SLAM robustness and hardware integration, the proposed framework provides a viable path toward reducing clinical staff burden and infection risks in future outbreak responses. The work is presented as a foundational step, with future iterations planned to integrate more robust SLAM methods and dedicated embedded hardware.

Technical Summary: Layered Risk Mapping for Autonomous Patient Transport in Expeditionary Medical Facilities

More like this