A Bayesian Reasoning Framework for Robotic Systems in Autonomous Casualty Triage

Szymon Rusiecki, Cecilia Morales, Pia Störy, Kimberly Elenberg, Leonard Weiss, Artur DubrawskiarXiv 2026 · arXiv

TL;DR

This paper introduces a neuro-symbolic cognitive architecture that combines robotic vision algorithms with an expert-elicited Bayesian Network to perform autonomous casualty triage in mass casualty incidents. By probabilistically fusing noisy, incomplete, and conflicting sensor data, the system achieved a nearly four-fold increase in overall triage accuracy during real-world field tests for the DARPA Triage Challenge.

What Problem Does This Paper Address?

In Mass Casualty Incidents (MCIs)—such as natural disasters, terrorist attacks, or armed conflicts—emergency medical services are instantly overwhelmed. While autonomous robots (UAVs and UGVs) offer a compelling solution for surveying disaster zones and remotely assessing injuries, real-world deployment introduces severe operational challenges. Dust, smoke, debris, and physical occlusions lead to noisy, fragmented, and incomplete sensor data.

Current robotic perception pipelines typically suffer from three critical flaws:

  1. Algorithmic isolation: Treating each vital sign or injury detector independently.
  2. Context blindness: Ignoring clinical inter-dependencies (e.g., how an amputation relates to blood loss).
  3. Brittleness: Underperforming or failing entirely under noisy, real-world conditions.

For instance, a standard robot might detect blood and a lower-extremity wound but fail to infer that the casualty is actively hemorrhaging—a critical insight needed to assign a high triage priority.

Methods

The authors propose a cognitive architecture centered on a Bayesian Network (BN) to act as a reasoning engine that integrates outputs from multiple independent perception algorithms into a unified probabilistic estimate.

  • Robotic Fleet: The system was deployed on a heterogeneous team of robots operating within a common ROS 2 framework, including three quadrupedal ground robots (Boston Dynamics Spot) and two Unmanned Aerial Vehicles (UAVs).
  • Perception Pipeline: Independent AI models process visual, thermal, and audio feeds to estimate specific signs like trauma location (head, torso, extremities) and alertness levels (ocular, verbal, motor).
  • Expert Knowledge Elicitation: Due to a lack of empirical training data, the structure and Conditional Probability Tables (CPTs) of the BN were engineered through an iterative process with three domain experts, each possessing over a decade of experience in emergency medicine. This clinical knowledge enforces realistic medical constraints on the robot's predictions.
The expert-elicited Bayesian Network architecture used to model clinical dependencies between vital signs and trauma indicators.
  • Probabilistic Inference: When a sensor fails or a casualty is partially hidden, the BN marginalizes over the missing variables. This allows the system to degrade gracefully and maintain real-time performance (<1 ms inference time, <100 MB memory footprint).

Datasets

Because no public, standardized datasets exist for robotic triage in MCI scenarios, a data-driven training approach was impossible. Instead, the framework was evaluated in realistic field tests constructed for the DARPA Triage Challenge (DTC) Round 1 Systems Competition.

The evaluation comprised 20 distinct casualty cases using a mix of human volunteers and high-fidelity TOMManikin trauma simulators distributed across unstructured terrain with varied injuries, postures, and environmental degradation (fog, dust).

Key Insights and Ideas

  • Neuro-Symbolic Synergy: Combining "black-box" neural network detectors with a "white-box" symbolic reasoning engine (the BN) provides both robustness to noise and clinical interpretability.
  • Graceful Degradation via Marginalization: Instead of failing when data is missing due to occlusions, the BN mathematically infers unobserved variables, maximizing the utility of whatever partial data is available.
  • Contradiction Resolution: The network resolves conflicting sensor inputs by enforcing clinical constraints. If a vision module incorrectly reports normal motor alertness due to minor movement artifacts, but another detects a catastrophic limb amputation, the BN recognizes the medical implausibility and prioritizes the severe hemorrhage.

Key Results

  • Massive Accuracy Boost: Overall triage performance (the rate of correct assignments across all vitals) increased nearly four-fold, jumping from 14% to 53% compared to a vision-only baseline.
  • High Reliability: The system's reliability—its ability to provide an assessment even with partial data—skyrocketed from 0.31 to 0.95, meaning it successfully generated triage profiles for 95% of the casualties.
  • Field Success: In the DARPA Triage Challenge, the framework more than doubled the baseline score in Run A (25 to 61 points) and nearly tripled it in Run B (16 to 45 points), securing 4th place out of 11 competing teams.

Limitations and Future Work

  • Cascading Error Propagation: Multi-modal architectures remain susceptible to cascading errors if an input module provides a highly confident but incorrect prediction.
  • Heuristic Calibration: The CPTs within the Bayesian Network were heuristically assigned based on qualitative expert statements (e.g., "highly likely"). Future work will focus on calibrating these probabilities as empirical data becomes available.
  • Human-Robot Teaming: For real-world deployment, future research must design intuitive interfaces that clearly communicate these probabilistic assessments and their associated uncertainties to human first responders.

Summary updated

Abstract

Autonomous robots deployed in mass casualty incidents (MCI) face the challenge of making critical decisions based on incomplete and noisy perceptual data. We present an autonomous robotic system for casualty assessment that fuses outputs from multiple vision-based algorithms, estimating signs of severe hemorrhage, visible trauma, or physical alertness, into a coherent triage assessment. At the core of our system is a Bayesian network, constructed from expert-defined rules, which enables probabilistic reasoning about a casualty's condition even with missing or conflicting sensory inputs. The system, evaluated during the DARPA Triage Challenge (DTC) in realistic MCI scenarios involving 11 and 9 casualties, demonstrated a nearly three-fold improvement in physiological assessment accuracy (from 15\% to 42\% and 19\% to 46\%) compared to a vision-only baseline. More importantly, overall triage accuracy increased from 14\% to 53\%, while the diagnostic coverage of the system expanded from 31\% to 95\% of cases. These results demonstrate that integrating expert-guided probabilistic reasoning with advanced vision-based sensing can significantly enhance the reliability and decision-making capabilities of autonomous systems in critical real-world applications.
BibTeX
@misc{rusiecki2026bayesian,
  author = {Szymon Rusiecki and Cecilia Morales and Pia Störy and Kimberly Elenberg and Leonard Weiss and Artur Dubrawski},
  title = {{A Bayesian Reasoning Framework for Robotic Systems in Autonomous Casualty Triage}},
  year = {2026},
  eprint = {2604.21568},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.21568},
}

Loading related content...

Figure