The thesis: one core, two markets
It is common to assume that a civilian inspection drone and a combat interceptor drone are different machines with different internals. At the level of the airframe and payload, that is true. But at the level of the "brain" — the part that turns a video stream into a decision — both are built around the same technological core.
That core consists of two linked capabilities: onboard AI object classification (a neural network on an edge device recognizes exactly what is in frame) and autonomous guidance without satellite (the drone holds and tracks an object by video, independent of GPS/GLONASS). This pair is the dual-use core: by changing the training data and the final-action logic, you get either a civilian monitoring tool or an active Counter-UAS defense system.
For a customer, this means clear engineering continuity: the same proven edge AI for drones architecture scales from routine power-line inspection to intercepting an FPV drone over a protected site. We cover each branch on its own in our articles on edge AI for civilian drones and on kinetic interception versus electronic warfare. Here, we focus on the bridge between them.
How the unified AI core is built
The processing pipeline is identical for both configurations and fits into four steps:
- Camera. An onboard video camera captures the scene in the visible spectrum. This is not a thermal imager — it is ordinary video that the onboard neural network processes.
- Onboard AI. Video feeds a Raspberry Pi-class edge computer right on the drone. All computation runs locally, with no frames sent to the cloud.
- Object-type classification. The neural network determines the class of the object in frame. In the civilian version it is a structural defect, a person, a fire ignition point; in the defense version it is a quadcopter, a fixed-wing UAV, a bird, or a manned aircraft.
- Autonomous decision. Based on the classification, the module decides on board: log and transmit the data, or hold the visual lock and carry the guidance through to the target without any satellite.
The key point is that onboard object classification and the decision happen in one device, in real time. In the VOLKODAV defense branch a second loop runs first: an AI server at the command post classifies video from all observation posts within 25–30 seconds and separates a real threat from a bird or a manned aircraft, with a false-positive threshold of no more than 5%. The interceptor's onboard AI module then takes over only the final phase. In the civilian branch the ground server is optional — all analytics can run directly on the drone.
It is worth understanding that the onboard AI module is the most expensive and knowledge-intensive part of the whole system. Most of the engineering effort goes into tuning it: collecting and labeling data, training the network for specific object classes, optimizing the model for the limited compute of an edge device, and field-testing under real lighting and weather. That is precisely why reusing this module across missions yields fundamental savings: once the "camera → AI → classification → decision" pipeline is tuned, it becomes a platform onto which you only need to fit the required set of classes and the required final action.
One core, two training regimes. The "camera → AI → classification → decision" architecture does not change — what changes is the classes the network sees and the action it responds with.
Civilian mission: monitoring and search
In the civilian configuration, the edge core turns the drone into an autonomous analyst that does not merely record video but immediately understands what is in it. This removes hours of footage review from human operators and delivers a result right in the field.
- Infrastructure inspection — power lines, comms towers, bridges, roofs, solar and wind farms. The onboard AI flags anomalies and defects during the flyover.
- Search and rescue — detecting a human figure in forest or field, including with a thermal payload.
- Forest monitoring — early detection of fire ignition points and illegal logging.
- Agriculture monitoring — crop condition assessment, field analysis, livestock control.
- Leak detection — thermal and visual anomalies on civilian power lines and pipelines.
What all scenarios share is that the decision is made on board, with no reliable link and no cloud. This is critical for remote sites and zones without coverage. A detailed breakdown of civilian applications is in our article "Edge AI on Raspberry Pi for Drones: Inspection, Search and Rescue, Forest and Power-Line Monitoring." We describe civilian applications as capabilities of the technology — honestly, without promising any specific outcome.
Defense mission: Counter-UAS and EW resilience
In the defense configuration, the same core drives an interceptor drone. The "camera → AI → classification → decision" logic stays, but the final action is a kinetic interception: an AI-guided interceptor physically rams the target in the air. With no explosives.
Two properties define this branch:
- Human in the loop at three independent points. The operator confirms detection, the pilot flies the drone over the radio channel, the operator confirms the visual lock — and only in the final phase does the autonomous onboard AI take over.
- Independence from GPS/GLONASS in the combat phase. Final guidance runs on the onboard camera video. Satellite jamming by electronic warfare has no effect — the drone sees the target visually and holds it autonomously. This is exactly why VOLKODAV is resilient to electronic warfare.
The interceptor is a quadcopter on 10″ propellers, up to 3 kg takeoff weight, with a cruise speed of 70 km/h (up to 120 km/h), a range of 10–12 km from the launch point, and up to 15 minutes airborne. Why kinetics are more EW-resilient than jamming, and when to choose which, is covered in our article "Kinetic Drone Interception vs Electronic Warfare" and in the landing-page section on EW resilience.
What is shared and what differs
To make the dual-use drone AI thesis concrete, let us put the configurations in a table. On the left is the component; on the right, how it looks in each version.
| Component | Civilian version | Counter-UAS version |
|---|---|---|
| Compute (edge) | Raspberry Pi-class, on board | Raspberry Pi-class, on board |
| Sensor | Video camera (opt. thermal) | Video camera |
| Classification network | Defects, people, fire, crops | Quadcopter, fixed-wing UAV, bird, manned aircraft |
| Guidance without GPS | Object tracking by video | Final target lock by video |
| Final action | Report, tagging, data transfer | Kinetic interception (ram, no explosives) |
| Ground loop | Optional | AI server at command post + human in the loop ×3 |
| Connectivity | Can operate without coverage | Pilot radio channel; final phase autonomous |
Three of the four core elements — the compute, the sensor, and the video-based guidance principle — are shared. What differs is the network's training classes and the final-action logic. This is the engineering essence of dual-use: the most expensive and complex part (the edge-compute and autonomous-vision architecture) is reused, while only the most application-specific part is adapted.
The difference in the ground loop is a separate, important detail. A civilian drone needs no ground infrastructure: it takes off, flies over the site, and returns with already-tagged data. Counter-UAS, by contrast, requires a full system — a command post with an AI server, automated observation posts with all-round thermal imagers, rapid-response teams, and a human in the loop at three points. This is not a contradiction but a reflection of a different cost of error: a false defect flag on an inspection is corrected by a re-survey, whereas an interception decision must pass independent human review at every step. The shared AI core removes the technical complexity of recognition, but the operational protocol of the two branches is fundamentally different.
Why the edge approach is the key advantage
The alternative to edge is to stream video to the cloud, compute there, and return the result. For both missions this is a losing scheme, and for the same reasons.
- Real time. The decision is made on board, with no delay for frame transfer and server response. For inspection that means a result during the flyover; for interception, reaction at a tempo a cloud round-trip cannot provide.
- Operation without a link. A remote pipeline, a forest without coverage, an electronic-warfare zone — the edge core does not depend on an external network. For a civilian drone this gives autonomy; for an interceptor, immunity to channel jamming.
- EW resilience. Because final guidance is video-based and computation is local, neither GPS jamming nor channel suppression deprives the drone of its ability to see and decide. This is a direct consequence of the edge architecture.
- Privacy and data control. Frames do not leave the device unless necessary — in civilian scenarios this simplifies work with sensitive sites.
In other words, edge is not an optimization but the very property that makes one and the same autonomous drone without GPS suitable for both monitoring and interception. For more on the system architecture, see the "How it works" and "Specs" sections on the home page.
Conclusion
VOLKODAV is not two different products but one edge AI core in two configurations. Onboard AI target classification and autonomous guidance without satellite equally underpin civilian inspection and defense Counter-UAS. The payload, the training classes, and the final action change — the core stays unified. For a customer this means a mature, reusable technology: what is proven on infrastructure monitoring strengthens defense against UAS, and vice versa.