Russian edge AI: the computers inside Russia's drones and missiles

Where Russia has put onboard computers that see — Nvidia Jetson modules, Raspberry Pi boards and optical navigation units — in its attack drones, loitering munitions and newest cruise missile, what each is said to do, how sure anyone is, and how the parts keep reaching Russia.

Updated

Since 2022 Ukrainian investigators have kept finding the same kind of part in Russian weapons: a small commercial computer, often an Nvidia Jetson module, wired to a camera. Ukraine’s defence intelligence (HUR) catalogues these parts in its War & Sanctions database. A Jetson TX2 turned up in a ZALA 421-16E reconnaissance drone captured in April 2022 and in the Lancet loitering munition. Jetson Orin modules followed in the V2U strike drone and an “MS series” Shahed in June 2025, and in the S-71M “Monokhrom” cruise missile in August 2026.

The hardware is the solid part of the story. What the software does is mostly known from what officials and analysts say about it. HUR says the V2U can find and pick targets on its own, and that the MS-series Shahed compares camera images with loaded models. Serhii “Flash” Beskrestnov, adviser to Ukraine’s president on defence technology, says Russian drones run the YOLOv5 object detector on Jetson boards. In August 2026 The New York Times reported that a Molniya with no radio link had picked its own final target in a strike that killed three civilians in Zaporizhzhia.

The applications are narrow: recognising objects, holding a lock in the final dive, and navigating by matching camera images to stored ones when satellite signals are jammed. A CSIS study by Kateryna Bondar (April 2026) found that more than half of 705 AI-relevant components recovered from Russian unmanned systems came from US-headquartered firms. On these pages, a hardware find is kept apart from a claim about what that hardware does.

Every application and platform carries an evidence level:

  • Confirmed teardown or independent documentation
  • Official claim a government or manufacturer says so
  • Claimed channels or press report it, unverified
  • Inference analysts deduce it from other facts

Which system does what

Each row is a system, each column an onboard function. A marked cell shows how well that use is documented: never better than the evidence for the system or for the function itself.

SystemTarget recognition and autonomous target selectionLock-on and terminal homingVisual navigation and terrain matchingStrike without a radio linkCamera guidance in a cruise missileGroup behaviour and swarmingRecognition on reconnaissance drones
V2UOfficial claimnot reportedOfficial claimOfficial claimnot reportedClaimednot reported
Geran-2 / Shahed-136 MS series (MS001)Official claimnot reportednot reportednot reportednot reportednot reportednot reported
Geran-4 (Seeker and Susanin versions)not reportedClaimedOfficial claimnot reportednot reportednot reportednot reported
Geran-5 K5not reportednot reportedOfficial claimnot reportednot reportednot reportednot reported
Molniya-2Official claimClaimedOfficial claimOfficial claimnot reportednot reportednot reported
S-71M "Monokhrom"not reportednot reportednot reportednot reportedConfirmednot reportednot reported
Lancet (Izdeliye-51, IRRA-equipped versions)Official claimClaimednot reportednot reportednot reportednot reportednot reported
Lancet-3 / Izdeliye-53not reportednot reportednot reportednot reportednot reportedClaimednot reported
ZALA 421-16Enot reportednot reportednot reportednot reportednot reportednot reportedClaimed
KUB-BLAClaimednot reportednot reportednot reportednot reportednot reportednot reported
Scalpelnot reportednot reportednot reportednot reportednot reportednot reportednot reported
Supercam S350not reportednot reportednot reportednot reportednot reportednot reportedClaimed
FPV strike drones (machine-vision versions)not reportedClaimednot reportednot reportednot reportednot reportednot reported
  • V2U

    Official claim
    Maker:
    Not established (HUR names two Russian and two Chinese firms in supply and assembly)
    Hardware:
    NVIDIA Jetson Orin Developer Kit on a Leetop a603 carrier board (Leetop A203 in early reporting), Sony IMX477 sensor, Meskernel rangefinder, KingSpec SSD, MicoAir M10G-5883 GPS, Microdrive LTE modem on early units
    Reported functions:
    target recognition; autonomous target selection; visual navigation; group flight (reported)
    First reported:
    2025-06-09

    source 1, source 2, source 3, source 4

  • Maker:
    Alabuga (Geran-2); HUR judged the MS001 airframe Iranian-built
    Hardware:
    NVIDIA Jetson Orin Nano Developer Kit, analog infrared camera via EZCAP USB 3.0 capture unit, STM32F103 "transition board for Nvidia Jetson", VHF datalink
    Reported functions:
    target recognition; autonomous target selection
    First reported:
    2025-06-27

    source 1, source 2, source 3

  • Maker:
    Alabuga
    Hardware:
    Raspberry Pi based Seeker kit (Beskrestnov; also fits Geran-2 and Geran-3); Susanin optical correlation unit and magnetometer unit in HUR's October 2026 record
    Reported functions:
    lock-on and terminal homing (claimed); visual navigation
    First reported:
    2026-06-30

    source 1, source 2, source 3, source 4

  • Geran-5 K5

    Official claim
    Maker:
    Alabuga
    Hardware:
    Susanin optical correction (OEKS / DSMAC class) on a unit labelled "Блок FCU Г5 K5-K Сусанин"; not yet in HUR's Geran-5 component list
    Reported functions:
    visual navigation
    First reported:
    2026-09-27

    source 1, source 2, source 3, source 4

  • Molniya-2

    Official claim
    Maker:
    Atlant Aero
    Hardware:
    NVIDIA Jetson Orin minicomputer (NYT, model not published); Jetson Nano with YOLOv5 (Beskrestnov); SIYI AI Tracking Module II (reported); Molniya-2R reconnaissance version with Raspberry Pi 5 and a Mini PC F8
    Reported functions:
    target recognition; terminal homing; visual navigation; strike without radio link
    First reported:
    2026-02-27

    source 1, source 2, source 3, source 4

  • Maker:
    Not confirmed (HUR names PJSC UAC as maker and designer of the related S-71K)
    Hardware:
    NVIDIA Jetson Orin TE980M-A1 (read as Orin NX 16 GB), production week 10 of 2025
    Reported functions:
    camera terminal guidance (claimed)
    First reported:
    2026-08-12

    source 1, source 2, source 3, source 4

  • Maker:
    ZALA Aero
    Hardware:
    NVIDIA Jetson TX2 180-83310-DAAF-C04; video processing unit, camera board, WONWOO video processor and Transcend NVMe SSD in HUR's March 2026 record
    Reported functions:
    image processing and target acquisition; terminal homing; target recognition (claimed)
    First reported:
    2023-12-18

    source 1, source 2, source 3, source 4

  • Maker:
    ZALA Aero
    Hardware:
    NVIDIA Jetson TX2 (ISIS teardown review)
    Reported functions:
    image processing and target acquisition; target-type discrimination (Russian claim); swarm testing (Russian claim)
    First reported:
    2023-10-28

    source 1, source 2

  • ZALA 421-16E

    Confirmed
    Maker:
    ZALA Aero
    Hardware:
    NVIDIA Jetson TX2 module with a grouped STM32F103C8T6 microcontroller, in a unit captured in April 2022
    Reported functions:
    onboard image processing (function not stated by HUR)
    First reported:
    2024-06-01

    source 1, source 2

  • KUB-BLA

    Claimed
    Maker:
    ZALA Aero
    Hardware:
    Not published
    Reported functions:
    target recognition (marketing claim; disputed)
    First reported:
    2022-05-26

    source

  • Scalpel

    Inference
    Maker:
    Vostok Design Bureau
    Hardware:
    Orange Pi single-board computer (Shenzhen Xunlong) and STM32H743 microcontroller in HUR's record; HUR says it is operator-guided and does not claim onboard recognition
    First reported:
    2026-03-23

    source 1, source 2

  • Maker:
    Unmanned Systems group (Supercam)
    Hardware:
    Not published
    Reported functions:
    object detection and classification in footage (claimed)
    First reported:
    2025-11-18

    source

  • FPV strike drones (machine-vision versions)

    Claimed
    Maker:
    Various
    Hardware:
    Not published
    Reported functions:
    automatic target acquisition and lock-on
    First reported:
    2024-01-06

    source

Target recognition and autonomous target selection

Official claim
A camera feeds an onboard computer that detects and classifies objects. HUR’s V2U record says the drone can detect and recognise objects and lock onto a target automatically (HUR). Of the MS-series Shahed, HUR said in June 2025 that its Jetson compares camera images with loaded models to pick targets (ArmyInform). In a September 2026 interview, Beskrestnov said data he had seen from Russian drones lists the classes their networks are trained on: tanks, vehicles, railways, locomotives, warehouses and a “soldier with a weapon” (DOU). The hardware is confirmed in wreckage. How well the recognition works has not been shown publicly.

Systems

  • V2UHUR says it can search for and select targets on its own; Jetson Orin on a Leetop carrier board source
  • Geran-2 / Shahed-136 MS series (MS001)Infrared camera and Jetson Orin; HUR says it compares imagery with loaded models to choose targets source
  • Molniya-2Beskrestnov says its network recognises tanks, vehicles, railways and locomotives source
  • Lancet (IRRA)ZALA promotes its IRRA recognition system; Russian channels credit it with finding camouflaged and moving targets (unverified) source
  • KUB-BLAZALA marketed AI target recognition; a CSIS review in May 2022 found the claim weakly supported source

Lock-on and terminal homing

Claimed
The most common use is the least ambitious. An operator picks the target, and the onboard computer holds the lock through the final dive, when the radio link often drops or is jammed. Militarnyi reported that the Lancet’s link tends to fail in the last metres, so the dive is flown on the optical sensor alone (Militarnyi). Beskrestnov said in June 2026 that the “Seeker” kit for Geran-2 to Geran-4 does recognition, lock-on and homing on a Raspberry Pi, so that a low-power jammer on the target will not stop it (Telegram). CSIS describes Molniya terminal guidance as following visual salience and motion contrast rather than classifying objects, and as continuing when the link is jammed (CSIS). In January 2024 Beskrestnov reported intercepting video from a Russian FPV drone with machine vision and automatic target acquisition (Euromaidan Press).

Systems

  • Geran-4 "Seeker"Several cameras, mesh-modem control, Raspberry Pi recognition and homing (Beskrestnov); redesigned nose with an EO sensor to finish the dive if the link is lost (Militarnyi) source
  • LancetOperator-guided; final metres flown on the optical sensor when the link drops source
  • Molniya-2Unfamiliar computer found between camera and mesh modem, reported as a SIYI AI Tracking Module II (February 2026) source
  • FPV strike dronesIntercepted video of a Russian FPV with machine vision and automatic target acquisition (Beskrestnov, January 2024) source

Hardware

Visual navigation and terrain matching

Official claim
Matching camera images against stored reference imagery lets a drone keep its course when satellite navigation is jammed. On 5 October 2026 HUR described an “electro-optical correlation system” called Susanin on an updated Geran-4 and said a similar one is used on the Geran-5 K5 (Interfax-Ukraine). Beskrestnov wrote on 10 October that Russia already uses similar navigation on the Molniya, V2U and Lancet (Telegram). For the V2U, HUR’s single GPS module led reporters to infer camera-based navigation in June 2025 (Kyiv Post). CSIS instead describes a laser altimeter and terrain-referenced flight profiles. In the Zaporizhzhia Molniya, Ukrainian officials told The New York Times that the unencrypted module held terrain images the drone could use to stay on course (Meduza, relaying the NYT).

Systems

Strike without a radio link

Official claim
The step after terminal homing is a drone that has no control link at all. In July 2026 Beskrestnov reported Molniyas carrying only a camera and a computer, with no antenna, in use in numbers around Zaporizhzhia, and warned that “a UAV detector no longer saves you” (Ukrainska Pravda). The New York Times reported that one such drone, flying in a group of about six with none transmitting, killed three people at a petrol station in Zaporizhzhia on 6 July 2026. Ukrainian officials said its code showed the target types it was trained on, and that the operator gave the area while the drone chose the final target (Meduza). CSIS reports that V2U units examined in mid-2025 carried LTE modems and that units intercepted in October–November 2025 did not (CSIS).

Systems

Hardware

Camera guidance in a cruise missile

Confirmed
On 12 August 2026 HUR reported a Jetson Orin module in the S-71M “Monokhrom” air-launched cruise missile and said only that it “may indicate” the use of artificial intelligence (HUR). HUR’s record dates the module to the tenth week of 2025 (HUR). The Register read the marking TE980M-A1 as a Jetson Orin NX 16 GB (The Register). That the module works with a camera for terminal guidance comes from the Ukrainian monitoring channel Colonel GSh, relayed by UNITED24 Media. HUR has not said so.

Systems

Group behaviour and swarming

Claimed
Claims of coordinated groups are the weakest part of the record. ISW reported in October 2023, citing Russian sources, that the Lancet variant Izdeliye-53 could tell target types apart and was being tested for synchronised swarm attacks (Espreso). CSIS reports that V2Us fly in staggered groups of six or seven. It calls the idea that wing markings serve as visual identifiers an inference, and it describes a reported May 2025 incident in which seven V2Us left their plan, circled and then attacked (CSIS). CSIS also finds no verified fielding of Kronstadt’s long-promoted Grom swarm concept.

Systems

  • Lancet Izdeliye-53Target-type discrimination and swarm testing, per Russian sources cited by ISW (2023) source
  • V2UGroups of six or seven; one reported coordinated attack in May 2025 source

Recognition on reconnaissance drones

Claimed
The same techniques sort reconnaissance imagery. HUR lists a Jetson TX2 module in a ZALA 421-16E captured in April 2022 (HUR). CSIS says ZALA has carried its civilian IRRA recognition software onto military platforms, including the ZALA T-16, T-20 and ZARYA. A spokesperson for Russia’s Unmanned Systems group told the state news agency TASS in November 2025 that the Supercam S350 was being integrated with software that detects and classifies objects in its footage (TASS). That is a manufacturer claim in state media, with no detail on whether the processing runs on board.

Systems

  • ZALA 421-16EJetson TX2 module listed by HUR; function not stated source
  • Supercam S350Object detection and classification software being integrated (TASS, state media claim) source

Where the chips come from

Nvidia stopped supplying Russia in 2022, yet HUR keeps finding its modules in new weapons. The Jetson Orin in the S-71M was made in the tenth week of 2025, three years after that exit (HUR; The Register). HUR does not say how any particular module reached Russia.

Chinese carrier boards. The Jetson in the V2U sits on a carrier board from the Chinese firm Leetoptech: an A203 in HUR’s first report (Kyiv Post) and an a603 in its component record (HUR). HUR said most of the V2U’s parts were foreign despite its Russian labelling, and named four firms in its supply and assembly, two Russian and two Chinese. The Molniya-2R’s mini-PC is a Chinese Mini PC F8 sold under the Russian Raskat label (Ukrinform).

Western parts, routed through third countries. CSIS counted 705 AI-relevant components recovered from Russian unmanned systems. More than half came from US-headquartered firms, which supplied about 69% of the memory, 57% of the processors and 38% of the sensors. China supplied under 9%. CSIS names China and Hong Kong as the main conduit after 2022, often through transshipment, and cites reporting of about $300 million in Nvidia chips shipped through Indian firms (CSIS). In March 2026 HUR counted 62 electronic components in the Lancet and Scalpel, mostly from the US, Switzerland and China (ArmyInform).

Sanctions response. In August 2025 Ukraine sanctioned 39 people and 55 companies in Russia, China and Belarus. They included ZALA Aero, Vostok Design Bureau and the AI drone centre Neurolab, along with ten Chinese component suppliers (Euromaidan Press). After the S-71M find, HUR said Russia still cannot replace foreign high-technology parts with its own and called for stronger, better coordinated sanctions (HUR).

Countering it

Defenders’ public statements agree on one point: jamming loses its value once the drone no longer needs a link.

  • Jamming works only before the lock. Ukrainian EW systems such as Bukovel and Nota were reported effective against the Lancet while it is operator-controlled, not after terminal lock (Militarnyi). Beskrestnov says a Seeker locked on a target is not stopped by a low-power jammer at the target (Telegram). CSIS notes that once the V2U dropped its LTE modem there was no signal left to jam (CSIS).
  • RF detectors miss silent drones. Beskrestnov’s warning about the no-antenna Molniya was that “a UAV detector no longer saves you” (Ukrainska Pravda).
  • Physical defeat. Ukraine’s air force spokesman Yurii Ihnat said in October 2026 that turrets aimed with machine vision had shot down jet drones around Kyiv, Geran-5s among them, at a range of about a kilometre (Ukrainska Pravda). Interceptor drones are the other answer Ukraine is scaling (Kyiv Independent).
  • Passive protection and decoys. Against the Lancet, nets and cage armour over guns and positions defeated impact-fuzed warheads until Russia moved to stand-off detonation. Decoys of artillery and launchers drew strikes away from real equipment (Militarnyi; Militarnyi).
  • Component denial. HUR publishes each module it finds so that partners can close supply routes (HUR).

Open questions

  • What software runs on the Jetson modules? No firmware dump or model file has been published. YOLOv5 is Beskrestnov’s account and CSIS’s reporting, not a published teardown.
  • Does the S-71M’s Jetson really guide the terminal phase, as a monitoring channel says? HUR has not described its role.
  • What does the Susanin unit contain, and is it a learned model or classic image correlation like the Kh-101’s DSMAC-class system?
  • How much of the Geran Seeker’s homing is autonomous? HUR does not use the Seeker name, and no teardown of the Raspberry Pi software is public.
  • Are the V2U group attacks reported by CSIS coordinated by software or by pre-planned routes?
  • How do Jetson modules and Leetop carrier boards reach Russian assemblers? No shipment-level evidence has been published for these parts.
  • Are fibre-optic FPV drones being fitted with terminal-guidance modules? No reliable public source documents it yet.
  • Is ZALA’s IRRA on the Lancet a version of its civilian software, and does it finish an attack after the link is lost, as Russian channels claim?

Sources

  1. Barrage munition with artificial intelligence V2U — components — HUR War & Sanctions 2025-06-09
  2. Ukraine Unpacks 'Russian' AI Drone: Foreign Tech, Chinese Parts — Kyiv Post 2025-06-09
  3. Computer module (NVIDIA Jetson Orin Nano Developer Kit) — Shahed-136 (Geran-2) MS001 — HUR War & Sanctions 2025-06-27
  4. With a camera and artificial intelligence: HUR on the upgraded Shahed downed over Sumy Oblast — ArmyInform 2025-06-27
  5. Latest Downed Shahed Drone – Hybrid Mix of Iranian, Russian Components — Kyiv Post 2025-06-27
  6. Computer module NVIDIA Jetson Orin TE980M-A1 — S-71M "Monochrome" — HUR War & Sanctions 2026-08-12
  7. War&Sanctions: Nvidia Jetson microcomputer found in new Russian Monokhrom missile — Defence Intelligence of Ukraine (HUR) 2026-08-12
  8. Nvidia Jetson module found in Russian S-71M missile — The Register 2026-08-14
  9. Russia's New S-71M Stealth Missile Can Hunt Targets With AI Powered by Nvidia Tech It Shouldn't Have — UNITED24 Media 2026-08-12
  10. Lancet loitering munition — components — HUR War & Sanctions 2026-03-23
  11. Scalpel loitering munition — components — HUR War & Sanctions 2026-03-23
  12. Ukraine's Military Intelligence Releases Detailed Information on Russian Loitering Munitions "Lancet" and "Scalpel" — ArmyInform 2026-03-23
  13. Russian Lancet-3 Kamikaze Drone Filled with Foreign Parts — Institute for Science and International Security 2023-12-18
  14. Jetson TX2 module — Zala 421-16E UAV (captured in April 2022) — HUR War & Sanctions 2024-06-01
  15. How Russia Is Building a Sovereign Drone Ecosystem for AI-Driven Autonomy — CSIS (Kateryna Bondar) 2026-04-13
  16. Russia Probably Has Not Used AI-Enabled Weapons in Ukraine, but That Could Change — CSIS (Gregory C. Allen) 2022-05-26
  17. Flash on jet drones and attacks on Kyiv (interview) — DOU 2026-09-04
  18. Russia fits Molniya strike drones with AI targeting modules — Hvylya 2026-02-27
  19. Russia begins mass deployment of autonomous AI-equipped drones in Zaporizhzhia — Ukrainska Pravda 2026-07-03
  20. Russian autonomous drone strike in Zaporizhzhia — The New York Times 2026-08-24
  21. NYT: A fully autonomous Russian AI drone run by an Nvidia minicomputer killed 3 civilians in Zaporizhzhia — Meduza 2026-08-24
  22. Telegram channel:Geran-4 Seeker has become common — Telegram — Serhii "Flash" Beskrestnov 2026-06-30
  23. Telegram channel:Flash: Susanin-based optical navigation found on Geran-4 and Geran-5 — Telegram — Serhii "Flash" Beskrestnov 2026-10-10
  24. Russia Upgrades Geran-4 Seeker Fuselage for Optical Guidance — Militarnyi
  25. Optical-electronic correlation system Susanin — Geran-4 (October 2026 version) — HUR War & Sanctions 2026-10-05
  26. HUR: updated Geran-4 with Susanin, MOK unit, KRSN and localised engine — Interfax-Ukraine 2026-10-05
  27. Telegram channel:Colonel GSh: Geran-5 K5 with OEKS (DSMAC) — Telegram — Colonel GSh 2026-09-27
  28. Russian forces begin using new Geran-5 version with autonomous optical navigation — Militarnyi 2026-09-27
  29. ZALA Lancet: engagement and countereffort — Militarnyi 2023-07-18
  30. The Lancet drone has acquired the ability to destroy targets protected by nets — Militarnyi 2023-11-09
  31. Russians deploy new attack drone with autonomous target identification capabilities — ISW — Espreso 2023-10-29
  32. Telegram channel, Russian side:IRRA-equipped Lancet strikes a moving tank (Russian MoD footage) — Telegram — Colonel Cassad (Russian milblogger) 2026-06-26
  33. Russian state media:Developers integrate Supercam S350 reconnaissance drone with artificial intelligence tools — TASS (Russian state media) 2025-11-18
  34. "They evolve": Ukraine tests FPV drone featuring target detection, lock-on, and tracking system — Euromaidan Press 2024-04-21
  35. Ukraine sanctions Russia's international AI drone networks — Euromaidan Press 2025-08-17
  36. DIU: Russia adapts Molniia drone for aerial reconnaissance — Ukrinform 2025-12-22
  37. Turrets have already downed jet-powered UAVs - Ukraine's Air Force spokesman — Ukrainska Pravda 2026-10-05
  38. Ukraine's interceptor tech outpaced by Russia's jet-powered Geran-5 drones — The Kyiv Independent 2026-10-01