field guide
Ukraine AI Terminal-Phase Drone Strikes
AI scoped to the terminal dive only - GPS-denied terrain matching, adversarial-hardened target recognition, and final-approach guidance - with the human in the loop for the rest of the mission.
Ukraine’s MoD integrates AI into “middle strike” drones for the terminal approach only — not the full mission. The AI handles three specific problems during the final dive: GPS-denied navigation via terrain matching, automated target recognition differentiating decoys from real equipment, and autonomous final trajectory adjustment. Operator control resumes only after the AI phase completes or is overridden.
This is a tightly scoped use of autonomy: AI is a terminal-phase specialist, not a mission manager. The rest of the flight is human-piloted with pre-planned routing that incorporates EW and radar coverage intelligence.
Architecture of the terminal AI stack
Three subsystems activate at the target grid boundary:
1. Visual terrain navigation (GPS-denied)
Onboard cameras scan terrain features — roads, rivers, landscape contours — and compare them against preloaded high-resolution satellite imagery. Position is determined by visual correlation, not GPS. This is critical because GPS is spoofed or jammed in the contested EW environment within 50-200 km of the front.
The problem reduces to image registration: a live camera frame matched against a geo-referenced satellite tile. Latency matters — the drone is moving at cruise speed toward the target, so the match must complete in a fraction of a second per frame for continuous position updates.
2. Automated target recognition (ATR)
Live video feeds are analyzed against trained visual patterns. The system discriminates between target types and assigns a tracking marker on valid lock. Multi-modal discrimination uses:
- Geometry: vehicle shape, dimensions, aspect ratio from approach angle
- Surface texture: paint schemes, material reflectance, thermal emissivity
- Thermal signatures: engine heat, exhaust plume, hot components
Russian countermeasure: zebra-stripe paint on vehicles designed to disrupt CV models trained on standard camouflage and vehicle shapes. This is an active adversarial ML problem — the defender (Russia) is generating perturbation patterns to fool the classifier without knowing the exact model weights. Effectiveness is unknown but reported as a concern.
3. Terminal guidance adjustment
Once locked, the AI controls the final dive trajectory in real time. This is a pure control problem — fast convergence to intercept, no overshoot, handling wind and target motion. The drone transitions from cruise (likely ~100-150 km/h for a fixed-wing chemical drone repurposed as a kamikaze platform) to terminal dive.
Pre-launch: route planning under EW constraints
AI is also used pre-launch to optimize flight paths. The planning system ingests:
- Radar coverage maps (known emitter locations, estimated detection ranges)
- Terrain masking opportunities (low-altitude corridors behind hills)
- Air defense blind spots (gaps in overlapping coverage)
- Target location and priority
The output is a waypoint route designed to minimize exposure to both detection and interception. This is classical operations research applied to contested airspace, not real-time autonomy — the route is computed once and uploaded before launch.
Logistics Lockdown program
The “middle strike” concept is part of a broader Logistics Lockdown program targeting Russian supply chains at 200 km depth. The operational logic:
- 50 km from front: targets are individual soldiers with backpacks (low cargo concentration, high protection density)
- 100-150 km: trucks and fuel tankers (medium concentration, lower protection)
- 200+ km: railheads, long-haul trailers, depots (highest concentration, lowest protection)
The deeper the strike, the more cargo density per drone sortie. This is a cost-efficiency argument: destroying one fuel tanker at 150 km removes multiple tons of fuel, equivalent to dozens of tactical strikes on individual vehicles near the front.
Key design constraint: terminal-only AI
The AI is deliberately NOT active during ingress or egress. This is not a technical limitation — it’s an operational choice:
- Reduces exposure to countermeasures (AI can’t be jammed/confused if it’s off)
- Keeps the human in control of routing decisions (terrain masking is hard to automate perfectly)
- Limits the autonomy surface area for safety and ethics
The drone is operator-flown until the target grid, at which point the AI takes over for the final approach. Operator retains override authority at all times — man-in-the-loop, not man-on-the-loop.
EW is not the whole counter-drone story
An important correction from operational feedback: electronic warfare accounts for only ~10% of counter-drone efforts. The rest is:
- Anti-drone nets strung along critical roads
- Air observation posts with shotguns every 50 meters
- Interceptor drones
- Physical hardening of vehicles and convoys
This matters because AI is often framed as the solution to EW jamming. It solves the GPS-denial problem, but the drone still has to physically reach the target through nets, shotguns, and interceptors. AI doesn’t help with physical interdiction.
See also
- ai-enabled-mid-range-strike-drones — the Azov unit-level perspective on the same campaign