Voron

Simulation work has a ceiling: at some point the drone has to be real. That's why I'm founding Artificial Intelligence in Aerospace, a new student organization at Georgia Tech launching in Fall 2026, with a team of 25+ engineers and one flagship program, the Voron drone project. The premise is to own the entire autonomy stack ourselves, from soldering the airframe to training the policy that flies it.

The core of Voron is a loop. We build the drones from hardware up, and alongside each physical airframe we build its digital twin, a simulation model matched to the real vehicle's mass, motors, and sensors. Policies train in the simulation environment through reinforcement learning, deploy to the physical drone, and every real flight generates telemetry that feeds back into the twin. That closes the gap between what the simulator believes and what the airframe actually does.

Perception comes from the drones' own eyes. Each vehicle carries visible-light video and infrared, and we'll train perception models on footage the fleet itself collects: EO/IR datasets of real environments rather than borrowed benchmarks, with IR extending operation into darkness and through conditions where video alone fails.

The connective tissue is the radio layer. Voron drones will talk to each other over a wireless mesh network rather than routing everything through a single ground link. Every vehicle is a relay, the topology heals itself when a node drops out, and coordination degrades gracefully instead of catastrophically. Getting RF communications right is unglamorous in exactly the way P&IDs are unglamorous, which is why I suspect most teams skip it and why we won't.

The simulation half of the loop already runs. Before the first airframe flies, I built the software backbone: a physics simulator that models the quadrotor down to its four motors, a photoreal Unreal Engine city to fly it through, and a PPO training pipeline where the learned policy is only allowed a bounded residual correction on top of a readable hand-written controller. Every checkpoint is graded on 100 fixed-seed episodes: with perfect state, with estimated state from noisy sensors, and with injected faults like GPS loss and motor degradation. That way, "it got better" is a measurement. The airframe above is the same model in all three places: the vehicle in the simulator, the asset in the renderer, and the reference the hardware team builds to.