Field Note · September 19, 2026
Field note: what breaks in the last 500 meters
Four failure modes that compound at precisely the moment a counter-UAS engagement is resolved — and why they are sensor architecture problems, not algorithm problems.
Most counter-UAS sensing discussions focus on detection range. That is the wrong end of the problem to worry about. Detection is largely solved by putting enough aperture and enough patience on the sky. The part that fails is the end of the engagement, and it fails for reasons that are architectural rather than algorithmic.
1. Blur arrives exactly when precision is needed
A frame sensor integrates over its exposure window. At the start of an engagement the target subtends few pixels and moves slowly across the field, so integration is harmless. In the terminal phase the angular rate climbs sharply and that same integration smears the target into a streak. The guidance loop needs a centroid at exactly the moment the sensor stops producing one.
No amount of downstream processing recovers information the sensor averaged away.
2. Dynamic range is decided globally
Exposure and gain on a conventional imager are scene-wide decisions. A small dark airframe against a bright sky, or anywhere near a low sun, forces a choice: expose for the background and lose the target, or expose for the target and lose everything else. Both choices are made before the tracker sees the data.
3. The compute does not fit
Dense frames at a rate high enough to fix problem one imply a processing load that wants a GPU-class part. An expendable effector has neither the mass budget, the power budget, nor the unit cost headroom for one. Frame rate and platform feasibility pull against each other and one of them has to give.
4. The latency budget is spent before inference
Terminal-phase detect-to-command budgets are measured in single-digit milliseconds. A pipeline organized around frame cadence has already consumed a meaningful fraction of that budget on capture and transfer before any useful work begins.
Why this points at event sensing
Each of the four is a direct consequence of synchronous, global, dense sampling. An event-based imager removes all three of those properties at once: per-pixel asynchronous response eliminates the exposure window, per-pixel logarithmic response removes the global exposure decision, and sparse output collapses the data volume — which in turn collapses both the compute requirement and the transfer latency.
That is not a claim that event sensing is better at everything. It is weaker at appearance, which is why we field a dual-modality configuration when classification matters. It is a claim that the specific failure modes of the terminal phase map cleanly onto the specific properties of the sensor.