Carnegie Mellon University Robotics Institute

Cecilia
Morales, PhD

Applied AI & Robotics Research Scientist

I take AI systems from the lab to the real world — building autonomous robots that triage and treat critically injured patients in mass casualty events, validated not just in simulation but on real hardware with real clinical constraints.

DARPA Triage Challenge Carnegie Mellon Robotics

Medical force multipliers

My research centers on a single question: what if one responder could save many? In mass casualty events, the gap between patients who need immediate care and the responders available to provide it costs lives. I build the systems that close that gap.

At Carnegie Mellon's Robotics Institute, advised by Artur Dubrawski, I developed autonomous robots that can assess patient vitals, locate hemorrhage sites, and perform procedural interventions — without waiting for a human to be free. Critically, each system was taken beyond simulation into real-world deployment: validated on physical robots, tested in animal surgical models, and evaluated under the constraints of actual clinical environments.

My philosophy: AI should extend human expertise, not replace it. I focus on bridging the gap between laboratory breakthroughs and field-ready systems — designing around uncertainty quantification, equity across patient populations, and robust performance on real hardware under real conditions.

3.8x
Improvement in triage accuracy with Bayesian reasoning
12+
Benchmark methods outperformed by PRISM
0.78 bpm
Mean absolute error on PURE heart rate benchmark
In vivo
Animal validation on real surgical hardware
Supported by
Uber Presidential Fellowship DARPA HR00112420329 NSF 2427948 & 2406231 U.S. Dept. of Defense

Two platforms, one mission

My dissertation organizes around two autonomous systems, each addressing a critical phase of mass casualty response.

🤖

Autonomous Multi-Patient Triage

A robotic system that navigates to patients, estimates vital signs through non-contact sensing, and makes triage decisions using a Bayesian network with embedded clinical knowledge — all in real time, across multiple simultaneous casualties.

🩺

Autonomous Hemorrhage Control

An ultrasound-guided robotic system for femoral vessel cannulation — locating bleeding vessels, tracking tissue deformation, and guiding needle placement to control life-threatening hemorrhage without a trained clinician present.

Sensing · rPPG

PRISM

Training-Free Adaptive rPPG

A parameter-free algorithm for estimating heart rate from video in challenging outdoor environments. Introduces spectral concentration and total variation metrics for adaptive signal optimization. Outperforms 12 state-of-the-art methods across 4 datasets, including the novel NATURE outdoor benchmark I collected and released.

Achieves 0.78 bpm MAE on PURE — outperforms 12 methods including supervised approaches, with zero training data.
Python Signal Processing NATURE Dataset rPPG-Toolbox ICRA
Hemorrhage · Ultrasound

RESUS

Ultrasound Augmentation for Cannulation

Augmentation pipeline for ultrasound imagery that improves downstream vessel detection and needle guidance in low-quality, real-world scans. Designed for data-scarce surgical robotics settings with limited labeled training examples.

Statistically significant improvement over non-augmented baselines on in vivo animal data.
PyTorch Ultrasound Data Augmentation Surgical Robotics
Triage · Uncertainty

MSU-Net

Uncertainty Quantification for Triage

A segmentation network with first-class uncertainty quantification, enabling the triage system to communicate confidence alongside predictions. Critical for deployable AI in high-stakes medical decisions with limited training data.

27.7% improvement in Dice score over Monte Carlo U-Nets, with 75–175x faster uncertainty computation.
Bayesian Deep Learning Segmentation UQ PyTorch
Hemorrhage · Navigation

BIFURC

Vessel Bifurcation Detection

Automated detection of femoral vessel bifurcation landmarks in ultrasound imagery — a critical navigation prerequisite for autonomous cannulation. Validated on in vivo animal data under real surgical conditions.

Expert-level bifurcation detection — largest real-world validation of any ML approach for US-guided femoral cannulation.
Computer Vision Ultrasound In Vivo Validation ROS2
Triage · Decision-Making

Bayesian Triage Network

Clinical Knowledge-Embedded BN

A Bayesian network integrating real-time sensor streams from ROS2 nodes with embedded clinical triage protocols. Enables data-efficient decision-making by grounding probabilistic inference in structured medical domain knowledge.

Triage accuracy from 14% to 53%. Diagnostic coverage from 31% to 95%. Inference in <1ms on edge hardware.
Bayesian Networks ROS2 Clinical AI DARPA
Infrastructure · Visualization

Needle Tracking & 3D Viz

Real-Time Surgical Tool Tracking

Real-time needle tracking under tissue deformation, plus a 3D ultrasound visualization tool for intraoperative spatial awareness. Built for deployment on actual robotic hardware with hard latency constraints.

KF-inspired block achieves best precision and needle tip error across all encoder architectures tested.
ROS2 3D Reconstruction Real-Time Hardware

Peer-reviewed contributions

Selected publications from my doctoral research on autonomous medical robotics.

A Bayesian Reasoning Framework for Robotic Systems in Autonomous Casualty Triage
Rusiecki, Morales, Dubrawski
ICRA 2026
Multimodal Bayesian Network for Robust Assessment of Casualty in Autonomous Triage
Rusiecki, Morales, Dubrawski
NeurIPS 2025 Workshop — Structured Probabilistic Inference & Generative Modeling
PRISM: Training-Free Adaptive rPPG Estimation in Challenging Outdoor Environments
Morales, Dubrawski
Under Review
Automatic Cannulation of Femoral Vessels in a Porcine Shock Model
Zevallos, Morales, Orekhov, Rane, Gomez, Guyette, Pinsky, Galeotti, Dubrawski, Choset
Hamlyn Symposium on Medical Robotics 2025

Let's build something that matters

I completed my PhD at Carnegie Mellon University and am actively exploring research positions in industry — particularly at organizations working on healthcare AI, medical robotics, embodied AI, or safety-critical autonomous systems.

If you're working on hard problems that need to move from the lab to the real world — where the stakes are real and the systems must actually work — I'd love to talk.