FIU Applied Research Center
Jan 2024 - PresentMiami, FL, USA
Robotics Research Assistant, Perception & Systems Lead
DOE MSIPP, with Savannah River National Lab & Oak Ridge Y-12
- Cut real-time inference latency 22× over an unoptimized PyTorch GPU baseline (CUDA, quantization, TensorRT) for a DeepLabV3+/STEGO segmentation pipeline that achieved a maximum of 72% mIoU across 21 navigation-critical classes, deployed onboard for real-time field demonstrations alongside a Gaussian Process model on one Jetson Orin Nano.
- Designed a sensor-agnostic ROS 2 node architecture so new sensors integrate without touching downstream navigation, mapping, or visualization, then deployed across multiple robotic platforms.
- Built distributed ROS 2 components in C++/Python and designed the robot's networking infrastructure from scratch (IP addressing, low-level Linux, DDS tuning, traffic prioritization), owning bottleneck analysis across the sensor-to-inference-to-operator path to hold real-time latency under load.
- Built an online Gaussian Process node in GPyTorch (CUDA), fine-tuned to the live CO2 sensor response; exposes feed and predict ROS 2 services that return mean and uncertainty for adaptive sampling, running concurrently with the segmentation model on the Jetson.
- Fused Velodyne lidar with six depth cameras into a unified point cloud; implemented SLAM with Nav2 planning and open-frontier exploration of unknown environments.
