Jaime A. Romero

I'm a graduate student at the University of Pennsylvania GRASP Lab, pursuing my Master's in Robotics.

Currently working at Whoop as an ML engineer on their Sensor Intelligence Team. Previously, I worked at General Motors on autonomous vehicle fleet systems and electric vehicle components. I completed my B.S. in Mechanical Engineering with a minor in Robotics at Carnegie Mellon University.

Email  /  CV  /  LinkedIn  /  Github

Research Interests

I'm focused on building safe, intelligent robotic systems for everyday environments. My work combines:

• Deep Reinforcement Learning – Training robots to learn complex behaviors
• Machine Perception – Computer vision, SLAM, and 3D reconstruction
• Social Navigation – Robots that navigate safely around people
• Autonomous Systems – Drones and humanoid robots for household applications

Projects

An interactive mechanistic-interpretability report on persistence—how many times a model re-attempts a staked riddle ("solve this and you get a Ferrari / save 10 lives") before giving up—and whether the activation energy (CETT) of specific MLP neurons on the opening prompt predicts it. The report toggles the full analysis across three models, combining a linear-probe regression, an L1-logistic context-fill classifier, a preference-vs-persistence choice experiment, activation steering, and a raw-conversation viewer.

In gemma-2-2b the coupling is real: a sparse set of importance neurons predicts persistence at AUC 0.82, and amplifying those neurons ×3 causally roughly doubles it. In qwen2.5-14B and deepseek-r1-14B it does not replicate—deepseek scores R² −0.38 on regression and AUC 0.625 against a 0.628 random baseline. Yet deepseek is by far the most persistent behaviorally (median 51 attempts vs. 2 for gemma and 4 for qwen, with 32% filling the 32K context), consistent with a reasoning model whose persistence is driven by the <think> trace after the prompt—invisible to a single opening-prompt forward pass.

[ Interactive Report ]

Implemented high-performance autonomous racing agent using Proximal Policy Optimization (PPO). Designed and iteratively refined reward shaping strategies—balancing gradual ramp-ups for track progress, gate passage, action continuity, and collision avoidance throughout training. Explored state estimation and integrated recurrent neural networks (RNNs) to improve policy performance. Also experimented with Soft Actor-Critic entropy stabilization methods to smooth reward transitions. Through extensive tuning and evaluation, achieved consistent lap times of approximately 18 seconds.

[ Video Demo ]

Investigated the sim-to-real performance gap between a drone trained in Isaac Lab and its real-world deployment. Conducted system identification and flight testing to compare simulated and physical behavior, identifying a consistent altitude offset of approximately one foot between environments. Developed calibration and control adjustments to compensate for the discrepancy and improve flight accuracy.

[ Video Demo ]

Repurposed the FlexBot—a differential-drive industrial robot donated to GRASP by Berkshire Grey—as a general mobile platform for future robotics research and teaching at Penn. Documented hardware and electrical architecture, evaluated custom control software, removed application-specific hardware, and integrated new systems including LiDAR, an external computer, and an odometry pipeline. Integrated the completed system into ROS 2 for teleoperation, odometry, and SLAM, and produced documentation for future users and development.

[ Video Demo ] [ GitHub ]

Recreated McGeer's passive dynamic walker through simulation and physical prototyping. The simulation utilized a 3-link Ordinary Differential Equation (ODE) model, transitioning to a 2-link ODE with intermediate velocity calculations based on inelastic collision assumptions.

[ Video Demo ] [ Numerical Method Simulation ]

Trained cartpole on stacked gymnasium obs data using SAC. While training, collected images of env and paired them with the enviroment obs. Trained a Resnet to predict the stack of enviroment obs based on a stack of env images. Without finetuning our actor performed very well!

[ Video Demo ]

Developed a Unity-based VR system that replicates the HINTS neurological exam to evaluate eye movement for diagnostic purposes. Conducted market research and customer discovery through CMU I-Corps program to validate clinical applications.

Built an in-house mobile fleet management system reducing routing and monitoring task times by ~80% and increasing data diversity by 13%. Deployed an image classification model (97% sensitivity, 96% specificity) for obstruction detection, improving fleet collection efficiency by ~6,800 hours/year.

Trained YOLOv5 object recognition model for road markings and defects (potholes, manhole covers, traffic cones). Developed a self-labeling pipeline with human-in-the-loop supervision to create the largest open-source pothole dataset. Designed integrated system architecture for autonomous navigation and pothole filling.

Designed evolutionary-learning-based algorithms for agent demographic selection in crowd simulation. Modeled pedestrian movement patterns using statistical analysis to improve robot navigation in human environments.

Experience

Autonomous Vehicles Fleet Systems SWE
General Motors, Warren, MI | June 2024 – July 2025
Fleet management systems, image classification, sensor calibration automation

Hydraulic Controls Group Product Engineer
General Motors, Warren, MI | June 2023 – June 2024
Transmission component design, supplier management, cost mitigation ($32M saved)

Electric Motors Intern
General Motors, Warren, MI | Summer 2022
Testing procedures for slot insulation, quality assurance, supply chain management

Research Assistant
CMU BIG Robotics Lab | Summer 2022
Social navigation algorithms, pedestrian modeling

Education

University of Pennsylvania
M.S.E. in Robotics | Expected May 2027 | GPA: 4.0
GRASP Lab

Carnegie Mellon University
B.S. Mechanical Engineering, Minor in Robotics | May 2023 | GPA: 3.7

Relevant Coursework: Deep Reinforcement Learning, Graph Neural Networks, Linear Controls, Machine Learning, Machine Perception, Physical Intelligence

Technical Skills

Programming: Python, C++, MATLAB, SQL, Linux
Robotics & ML: ROS, YOLOv5, SLAM/VSLAM, PyTorch, TensorFlow, Gymnasium, Gazebo
CAD & Simulation: Siemens NX, SolidWorks, AutoCAD
Manufacturing: Ultrasonic Welding, Laser Welding, Wire Bonding, GD&T

Last updated: June 2026 | Design from Jon Barron's website