Field notes on CUAS, defense AI, 2POV drone-to-drone capture, and the operational realities of building models that perform against real adversary drones — not the ones in stock footage.
A Group 1 quadcopter and a Group 2 fixed-wing don't sign the same in the sensor stack. Why CUAS detectors need class-specific training data.
Read →Adversary drones don't stop flying at sunset. Thermal and low-light training data is the missing half of most CUAS detector pipelines.
Read →A T&E harness that can't be re-run against new adversary platforms produces performance numbers with a short shelf life. What a living harness needs.
Read →Swarm behavior is a distinct detection problem, not an extrapolation of single-target CUAS. What that means for training data.
Read →Synthetic drone data has a place in CUAS training pipelines. That place is not "instead of real data." Where each actually belongs.
Read →The adversary drone market is commoditizing faster than our defenses. A look at what training data the next generation of C-UAS needs.
Read →A dedicated camera operator tracking the target drone changes what a dataset can teach a model. Here's why drone-to-drone 2POV matters.
Read →Your detector is only as good as the distribution it was trained on. Most public datasets were built for a world that doesn't shoot back.
Read →Detectors don't fail on the average case. They fail on the weird one. Engineering a dataset that covers the long tail on purpose.
Read →Offshore annotation is cheap. Offshore capture is a non-starter for most defense and CUAS programs. Why the crew matters as much as the payload.
Read →Getting a detector trained is the easy part. Getting it fielded, monitored, and re-trained on program timelines is where most efforts die.
Read →Automatic Target Recognition lives and dies on class definitions. Get the ontology wrong and no amount of data will save the model.
Read →