denied.systems

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.

updated 2026-06-29 autonomy ukraine operational case-study

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:

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:

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:

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:

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:

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