Autonomous systems
UAH · NATO SPS
My work on SAPIENCE
I built the development and deployment workflow for our six-person student team, supported simulation at fleet scale, and helped fly our first flight on visual-inertial odometry.
AI engineer & pilot
I’m Joey Schwalb.
I’m a senior AI engineer at Camgian and a commercial pilot.
With my team at UAH, I built a software factory that took code from push to real hardware, through multi-architecture images, physics simulation and a synced ground station. From that station, I helped fly our first autonomous flight on visual-inertial odometry.
At Radiance Technologies, I was a major contributor to modeling and simulation platforms. Then, as an ML research scientist, I built the MLOps infrastructure and GPU clusters behind my reinforcement-learning work.
I also built a Unity research environment, used in more than 100 experiments, that became the foundation of my master’s thesis.
Based in Huntsville, Alabama.
Some of my work
A few projects from UAH and Radiance Technologies, with notes on what I built and how it worked.
Infrastructure
UAH · SAPIENCE
Deploying a fleet in disconnected environments
With my team, I built the tooling that stands up K3s, distributes images, and deploys services across the aircraft and ground station. We deployed three real aircraft with it, and scaled it to 25+ drones in simulation.
Flight
SAPIENCE · July 2026
Flying with visual-inertial odometry
In July 2026, during SAPIENCE development, we made our first autonomous flight with VIO as the estimator's only source of odometry. I worked on the robotics software and helped fly it from the ground station, deploying and monitoring the stack.
Simulation
UAH
Testing drones in simulation
I ran 25+ drones in PX4 SITL on consumer desktop hardware using K3s. I also built a Unity research environment used in more than 100 experiments.
Machine learning
Radiance · UAH
Training clusters for reinforcement learning
At Radiance Technologies, I built GPU training clusters and pipelines for reinforcement-learning control studies in AFSIM. Separately, at UAH, I contributed to training detectors and containerized them for deployment on embedded systems.
About me
I studied aeronautics at the University of North Dakota, then computer science at UAH, and earned a graduate certificate in cybersecurity from Harvard Extension School. I’m also a commercial pilot with instrument and multi-engine ratings.
Experience, education, and publicationsAs an Autonomous Systems Engineer at UAH, I worked with a six-person student team on the NATO-funded SAPIENCE project. I built the development and deployment tools we used to test changes and get software onto the aircraft.
As a Machine Learning Research Scientist at Radiance Technologies, I worked on reinforcement learning in AFSIM and the infrastructure for training those models. That included GPU clusters and pipelines built with Ray, MLflow, Docker, and GitLab.
Earlier, as an intern at Collins Aerospace, I built flight-data-recorder tooling and certification test cases for Bombardier avionics, and at UND I built UAV swarm platforms flown in test flights. I’ve published three IEEE papers on human-AI teaming and UTM safety, and I’ve worked in Python, C++, PyTorch, ROS 2 and Kubernetes.
I publish research as Joseph Schwalb. My M.S. thesis at UAH studied explainable AI and human-AI teaming in virtual environments.
I’m working toward a Ph.D. in Computer Science at UAH, with graduation expected in Fall 2027.
Try it yourself
Take an aircraft apart, replay a flight, follow one commit to a disconnected fleet, send a drone to a simulated survivor, or train 128 simulated vehicles. It all runs in your browser, with nothing to install.
Start with the aircraft
Loading the interactive route
This browser could not start the 3D view, so this is a still image of it.
Aircraft
The aircraft
At UAH, our SAPIENCE team built search-and-rescue drones from off-the-shelf and printed parts. I worked on robotics software, machine learning pipelines and fleet deployment.
Interactive illustration. The fleet and radio traffic are simulated. Right-drag or use two fingers to orbit.
Infrastructure
From a laptop to a launch
With my team, I built the SAPIENCE software factory: automated builds, simulation tests and airgapped deployment to our search-and-rescue (SAR) fleet.
- Write. Six of us worked on one autonomy stack.
- Push to the software factory. GitLab CI bakes only the changed images, for amd64 and arm64. I also worked on SBOMs and provenance.
- Test in Gazebo. On request, tagged images bootstrap Gazebo experiments: whole SAR missions in PX4 SITL. I built the contract tests and fuzzers, and a grader scored each run from its logs.
- Sync the bundle. One CLI stands up K3s and joins the nodes over a batman-adv mesh. While connected, the control plane copies each listed artifact, upstream dependencies included.
- Cross the airgap. With no reliable internet in the field, K3s’s built-in Spegel mirror shares images peer to peer. The ground station seeds them, and each drone that pulls one serves it to the rest.
- Deploy declaratively. One Helm package gives each drone its pod, on a Jetson or in simulation.
- Start the pods. For observability, Prometheus scrapes the services, and Grafana charts them.
- Report mission-ready. No drone leaves alone: each reports ready, then holds until it hears the others.
- Launch. With all three mission-ready, they lift off and the search starts.
Deployment case study · Simulation case study
Architecture illustration. A drawing of what we ran. A stripped-down stack was used for flight.
Mission
The mission
At the third NATO-funded SAPIENCE competition, our three aircraft had to fly a search-and-rescue scenario: locate eight survivors (mannequins) across a disaster zone and three collapsed buildings (tents) without GNSS, triage them, and land beside each one with the right package.
I built the deployment and simulation tests behind them, contributed to training the detectors, and containerized them for embedded deployment.
Mission illustration. Simulated in your browser from the competition’s rules, not a recording: no team served all eight in the real mission. Time is compressed and objects enlarged.
Flight
Flying with visual-inertial odometry
In July 2026, the stack from our software factory took off and flew our SAPIENCE aircraft with visual-inertial odometry (VIO) as its only odometry. I helped fly it from the ground station, deploying and monitoring the stack.
Getting the ZED SDK’s VIO working on our platform was non-trivial. Of all the SAPIENCE teams, our system flew the longest GNSS-denied, and it was the lightest, by almost half the weight.
- Deploy. Our factory’s images start on the Jetson Orin Nano, between the ZED 2i and PX4 on the Pixhawk 6X.
- Settle. On the ground, the VIO settles until EKF2 takes its positions. Simulated, with PX4’s default gate. Pick the ground:
- Take off. From here it’s the log. Each dot is a VIO position less where EKF2 predicted it, and through the climb every one fused. The loupe also shows the IMU and the motors.
- Fly. The stack flies the first two minutes itself, on VIO, then a pilot flies it in position mode and lands. The ghosts are where it believed it was. The line is where it went, from the GPS we logged.
These are real logs from our inagural VIO flight in July 2026!
Machine learning
Machine learning, at work and in your browser
At Radiance Technologies I developed online reinforcement-learning methods for AFSIM control studies, and built the GPU training clusters and pipelines behind them with Ray, MLflow, Docker and GitLab.
Browser demo. 128 simulated vehicles try to land inside a crater, learning live. Arrows show what the policy commands, and dark lines how it flies now. The alien ship is domain randomization: it keeps reshaping the field while they train.
Perception
Finding people from the air
At UAH, I contributed to training the detectors, then containerized them for deployment on embedded systems.
Browser demo. Pretrained public models, running in your browser. The photo or live stream stands in for what the ZED 2i sees, and the hover, desk and wall are simulated.
Get in touch
If you’re hiring or have a question about my work, I’d be happy to hear from you. I can send a résumé.
