DARPA Triage Challenge

• DARPA Triage Challenge is a multidisciplinary initiative that develops fully autonomous robotic systems for medical triage during Mass Casualty Incidents (MCIs).
• It integrates UAVs and UGVs equipped with multi-modal sensors and advanced perception algorithms to detect, assess, and locate casualties in complex, unstructured environments.
• Field evaluations show improved triage speed and diagnostic accuracy through innovative sensor fusion and Bayesian reasoning, paving the way for robust disaster response.

The DARPA Triage Challenge (DTC) is a multidisciplinary research initiative designed to advance the development and deployment of fully autonomous robotic systems for medical triage in Mass Casualty Incidents (MCIs). The core objective is to enable robots to locate, assess, and report critical physiologic signs and injuries for multiple casualties in complex, unstructured environments, without human intervention or physical contact. The DTC evaluates integrated robotic solutions across realistic field scenarios to benchmark their effectiveness and reliability in disaster response contexts (Rusiecki et al., 21 Dec 2025, Hughes et al., 9 Dec 2025).

1. Problem Scope and Evaluation Criteria

The DTC targets the automation of primary triage during MCIs—events where the number and severity of casualties overwhelm traditional emergency medical systems, causing critical delays and likely preventable mortality. Competing systems must autonomously perform the following functions for each casualty:

• Detection and Localization: Identify casualty locations in challenging environments (e.g., open battlefields, convoy ambushes) featuring occlusion, dust, rough terrain, mannequins, and live actors.
• Physiological and Injury Assessment: Non-contact determination of:
  • Severe Hemorrhage (present/absent)
  • Respiratory Distress (present/absent)
  • Head, Torso, Upper/Lower Extremity Trauma (categorical)
  • Ocular, Verbal, Motor Alertness (e.g., eye status, speech, movement)

Each scenario typically consists of up to 30-minute continuous runs. The DTC employs rigorous quantitative metrics:

• Physiological Assessment Accuracy: Fraction of vital sign measurements matching ground truth (when attempted).
• Triage Accuracy (Performance): Fraction of all required assessments correctly completed.
• Diagnostic Coverage (Reliability): Fraction of required fields actually attempted.

Scoring allocates a maximum of 12 points per casualty, with higher weight assigned to correct and timely identification of critical conditions during the "Golden Window" (first half of the trial), specifically: Severe Hemorrhage and Respiratory Distress (4 points each if early, 2 points otherwise), Head Trauma (2 points), and Torso Trauma (1 point) (Rusiecki et al., 21 Dec 2025).

2. Robotic Architectures and System Design

Solutions competing in the DTC utilize heterogeneous, modular robotics platforms for robust field triage. Prominent design patterns include coordinated unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs):

• Aerial Platform (UAV): Employs advanced imaging (RGB, LWIR/thermal cameras) for wide-area victim detection and geolocation, supported by GNSS, inertial measurement, and mesh networking. The Falcon 4 quadrotor, for example, incorporates custom carbon-fiber frames, NVIDIA Jetson compute, and multi-modal sensors for day/night operations (Hughes et al., 9 Dec 2025).
• Ground Platform (UGV): Outfitted with high-throughput computing hardware (e.g., AMD Ryzen CPUs, NVIDIA RTX GPUs), modular "plug-and-play" sensor racks, multi-modal perception (RGB, thermal, event, radar, LiDAR), and acoustic interfaces. These platforms execute close-range vital signs extraction, injury detection, and audio-based alertness probes.

System integration leverages mesh radio networks, distributed databases (MOCHA), and resilient communication stacks to maintain real-time data flow, even under degraded connectivity. Basestation components aggregate and visualize incident data, providing critical user interfaces for field operators and first responders (Hughes et al., 9 Dec 2025).

3. Sensing, Perception, and Data Fusion

DTC systems implement a repertoire of classical and machine learning algorithms for robust, multimodal casualty assessment. The perception stack supports:

• Victim Localization: UAVs deploy YOLOv8 models fine-tuned on in-house and public datasets for day/night detection, with reprojection from image to world coordinates for geolocation (mean error ≲2 m). UGVs refine locations by integrating instance segmentation and LiDAR/EKF fusion.
• Vital-Sign Measurement: Heart rate (HR) via remote photoplethysmography (rPPG) and radar, using signal processing pipelines (e.g., CHROM, multitask temporal-shift networks). Respiration rate (RR) from pulsed radar, LWIR ROI temperature traces, and chest motion event cameras.
• Injury and Status Classification:
  • Vision-LLMs (VLMs): LLaVA-8B and NVILA-Lite-2B for field-specific trauma and alertness assessment; fine-tuned via cross-entropy loss on multi-view vignettes.
  • Grounding DINO Pipeline: Person and injury detection (blood, wounds, amputations), body-part association, and semantic region assignment.
  • Specialized Classifiers (DINOv3): Binary classification for hemorrhage and trauma, with weighted loss functions for category imbalance.

Differentiated fusion strategies are used at the triage level. For instance, a Bayesian network constructed from domain-expert rules (not data-driven) can systematically combine noisy, incomplete outputs from multiple modalities. This approach supports reasoning in partially observable environments and demonstrates substantially improved diagnostic coverage compared to vision-only baselines (e.g., 95% vs. 31% case coverage, 53% overall triage accuracy vs. 14% for baselines) (Rusiecki et al., 21 Dec 2025).

Final triage decisions employ weighted score fusion and established triage protocols (e.g., SALT rules), considering uncertainty and variance of individual modality outputs (Hughes et al., 9 Dec 2025).

4. Coordination, Autonomy, and Human-Robot Teaming

Autonomous task execution in the DTC is governed by multi-layered coordination and planning schemes:

• Task Allocation: UGVs dynamically select and service the nearest untriaged casualty identifier, using geospatial assignments derived from UAV mapping, optimizing for minimal travel time.
• Path Planning: Global waypoints are generated using GNSS data, with LiDAR-powered local obstacle avoidance via sampling-based planners (e.g., RRT*).
• Autonomy Modes: Human supervisors may trigger specific measurement events ("collect vitals") in current deployments, with forward pathways aiming for full autonomy via traversability mapping and state-of-the-art distributed navigation.

Communication protocols combine mesh radios for low-latency telemetry and opportunistic data sync mechanisms (MOCHA) to ensure resilience. Operator interfaces include the Android Team Awareness Kit (ATAK) for live triage status and web UIs for detailed visual streams and geospatial visualization (Hughes et al., 9 Dec 2025).

5. Experimental Results and Benchmark Performance

DTC scenarios rigorously assess system performance in operationally realistic settings:

• Victim Detection: Mean average precision at IoU=0.5, 82% (day, RGB) and 76% (night, LWIR).
• Localization Accuracy: UAV global reprojection errors ≈1.8±0.9 m; UGV refinement to ≈0.8±0.4 m.
• Vital Signs Estimation: Heart-rate mean absolute error (mmWave) ≈4.6 bpm; respiration rate MAE (LWIR, pulsed radar) <2 bpm.
• Injury and Status Classification Accuracies (Year 2, held-out validation):

| Field | VLM (%) | DINOv3 (%) | Grounding DINO (%) | |-----------------------|---------|------------|--------------------| | Severe Hemorrhage | 47.8 | 50.5 | 43.9 | | Ocular Alertness | 80.6 | 56.9 | 56.8 | | Head Trauma | 79.1 | 69.1 | 83.5 | | Torso Trauma | 74.6 | 52.1 | 62.6 | | Lower Extremity Trauma| 64.2 | 35.6 | 48.2 | | Upper Extremity Trauma| 68.7 | 55.3 | 54.0 |

• Triage Throughput: Mean triage time reduced from 60±12 s/casualty (remote) to 45±8 s/casualty (onboard/automatic).
• Communication Reliability: >98% opportunistic data sync success rate under 2 Mbps mesh network conditions; ATAK interface latency 0.4±0.1 s.

A Bayesian network-based fusion strategy produced a dramatic improvement in both accuracy of physiological assessment (e.g., from 15% to 42% and from 19% to 46% in field trials) and diagnostic coverage (from 31% to 95%) compared to vision-only baselines, substantiating the impact of expert-guided probabilistic reasoning (Rusiecki et al., 21 Dec 2025).

6. Scientific Contributions and Limitations

The DTC has produced significant advances in autonomous triage, with several notable innovations:

• Integration of Classical and Foundation Models: Systems combine signal-processing algorithms with transformer-based VLMs and vision encoders (e.g., DINOv3).
• Expert-Knowledge-Guided Probabilistic Reasoning: Deployment of Bayesian networks constructed solely from expert rules, underlining the robustness and sample efficiency of domain-guided approaches in sparse, high-noise settings.
• Sensor Modularity and Orchestration: Platforms support rapidly reconfigurable, sensor-agnostic orchestrators for multi-modal fusion.
• Real-World Readiness: Solutions incorporate HDR imaging for challenging light conditions, real-time onboard inference, and communication schemes robust to unreliable field networks.

Observed limitations include persistent difficulty in robust hemorrhage classification (accuracy plateauing at ~50%), motivating further dataset diversification and exploration of synthetic data augmentation (e.g., Stable Diffusion-generated samples). Current deployments remain human-supervised for some triggers, with future directions targeting full robotic autonomy and uncertainty quantification in triage reports (Hughes et al., 9 Dec 2025).

7. Broader Implications and Future Trajectories

The DTC demonstrates that integrated, heterogeneous robotic systems—coupling classical signal processing, modern AI models, and expert-encoded reasoning—can meet key benchmarks for autonomous triage in MCIs. The results underscore the importance of modularity, domain knowledge integration, and system-level optimization for robust decision support in high-risk, operationally complex environments.

A plausible implication is that these architectures could serve as blueprints for broader deployment in disaster medicine, particularly where human responders face access barriers or information gaps. Expanded scenario diversity (night, occlusion, adverse weather), improved hemorrhage detection pipelines, and fully autonomous end-to-end systems are identified as near-term priorities. Embedding epistemic and aleatoric uncertainty into triage outputs is also a developing area, critical for downstream medical decision support (Hughes et al., 9 Dec 2025).