Work · US Navy research programme
Natural-Language Flight Control
A chat-to-fly interface for UAV operators: spoken or typed instructions translated directly into flight commands by a language model running on the aircraft, not in a datacentre.
The problem
Operating a UAV under load means translating intent into a sequence of low-level commands, which is slow and error-prone precisely when attention is scarce. The obvious fix, a language model, normally assumes a datacentre behind it.
This programme could assume neither connectivity nor power. The model had to run on the aircraft, inside a small power envelope, over a link that could not be trusted.
What I did
Fine-tuned BERT-based models to map natural-language instructions onto a constrained command vocabulary, so the output space was verifiable rather than open-ended.
Quantised to INT8 and profiled against the airframe power budget until inference fit on embedded hardware, running the pipeline across Jetson Orin, Raspberry Pi and RP2040 class devices.
Built the ground-to-air communications stack alongside it, with segmented networking and encrypted telemetry, then threat-modelled and penetration-tested the result rather than assuming it was secure.
Result
Operators could issue intent in plain language and have it executed as precise flight commands, with situational awareness returned in the same channel. The full stack, custom airframe, hardened comms and on-device NLP, ran without a cloud dependency.