AI drones explained: autonomy, guidance and human control

Ask which task is autonomous

Autonomy does not necessarily require machine learning. The ICRC notes that some autonomous systems use simple rule-based software. Conversely, a weapon using AI only for navigation is not necessarily an autonomous weapon in the targeting sense. An “AI” label cannot settle the classification.[6]

GNSS means global navigation satellite systems, including GPS. Navigation without those signals can draw on inertial sensors, visual odometry, terrain matching or simultaneous localisation and mapping. RUSI lists these alternatives. They are a family of techniques; identifying the technique matters more than the headline label.[9][10]

CSIS describes last-mile guidance as the final approach after an operator has visually confirmed a target. The operator activates lock-on; the onboard system then guides the approach. This can reduce reliance on a continuous manual command link during that phase without proving independent selection of a new target.[8]

Recognition and permission to attack should be read separately. Reporting on Saker Scout relays the Ukrainian ministry’s claim that the system detects equipment and sends coordinates to a command post. Detecting an object, classifying it as equipment and authorising force are not interchangeable descriptions.[6][11]

A sourced timeline

1. 2021-12-09

   DARPA’s OFFSET programme reports experiments in swarm autonomy and human–swarm teaming. Here, “swarming” is a question about coordination and allocation of tasks among vehicles. Footage showing many drones at once is insufficient to establish what coordination was automatic or who authorised their actions.[1]

2. 2023-01-25

   The January 2023 edition of US DoD Directive 3000.09 requires appropriate levels of human judgment over the use of force and sets review procedures, with specified exceptions and a waiver route. That dated US policy is not a global ban or proof that all autonomous engagement is prohibited.[2]

3. 2023-08-30

   The 2023 Nature paper on Swift demonstrates autonomous drone racing using learning-based and traditional algorithms. It is peer-reviewed evidence for a specific perception-and-control task. Winning a defined race is not evidence that the same system can distinguish lawful targets in an uncontrolled combat environment.[14]

4. 2023-09-04

   Ukrinform reported Ukraine’s approval of Saker Scout on 4 September 2023, citing the defence ministry. The claimed recognition and coordination functions are official claims relayed by the agency. Approval for use is not an independently published accuracy test across camouflage, weather and different target classes.[11]

5. 2025-03-06

   CSIS describes last-mile guidance as the final approach after an operator has visually confirmed a target. The operator activates lock-on; the onboard system then guides the approach. This can reduce reliance on a continuous manual command link during that phase without proving independent selection of a new target.[8]

6. 2026

   The UN Convention on Certain Conventional Weapons (CCW) process scheduled expert-group sessions on lethal autonomous weapons for March and August–September 2026. The ICRC’s 2026 FAQ distinguishes autonomous navigation from autonomous targeting. Negotiating sessions, proposed rules and existing obligations are different things; none is a performance test of a named drone.[6][16]

What is established, and what is claimed?

Named systems, carefully bounded evidence

Ukrinform reported Ukraine’s approval of Saker Scout on 4 September 2023, citing the defence ministry. The claimed recognition and coordination functions are official claims relayed by the agency. Approval for use is not an independently published accuracy test across camouflage, weather and different target classes.[11]

Lancet’s autonomy claims have a history. A May 2022 CSIS analysis distinguished Kalashnikov’s marketing of autonomous target-finding from evidence of actual use in that mode. It is a dated warning about proof standards, not a current verdict that later Lancet versions lack autonomous functions.[7]

Reporting that cites Ukrainian defence intelligence describes a recovered Shahed variant with an infrared camera and an Nvidia Jetson Orin computer. That supports a reported hardware change. The presence of an AI-capable processor does not, on its own, establish which software ran or whether targets were selected autonomously.[12]

AeroVironment says Switchblade 600 operators can wave off an attack and re-engage. That is a named manufacturer claim about an operator-control feature. The fact that a munition has onboard guidance does not tell you that human supervision or an abort option is absent.[15]

The US Army’s HAROP reference describes an operator selecting one loitering munition for target search and attack while monitoring others. This illustrates why the label “loitering munition” does not by itself specify who chooses a target or how much control is delegated to software.[4]

In a CSIS interview, Shield AI described V-BAT transitioning from GPS to an alternative navigation solution. Treat this as the manufacturer’s explanation of its system. It does not independently verify every advertised autonomy feature or establish autonomous authority to use force.[13]

A swarm claim needs more than a crowd

DARPA’s OFFSET programme reports experiments in swarm autonomy and human–swarm teaming. Here, “swarming” is a question about coordination and allocation of tasks among vehicles. Footage showing many drones at once is insufficient to establish what coordination was automatic or who authorised their actions.[1]

The 2023 Nature paper on Swift demonstrates autonomous drone racing using learning-based and traditional algorithms. It is peer-reviewed evidence for a specific perception-and-control task. Winning a defined race is not evidence that the same system can distinguish lawful targets in an uncontrolled combat environment.[14]

Failure modes and countermeasures

An August 2024 CSIS commentary relays Ukrainian accounts of vision-equipped US drones struggling over dense forests. This is a context-specific report of navigation failure. It is a reason to ask about realistic test conditions, rather than assuming a successful demonstration proves reliability in every landscape.[3]

The countermeasure question depends on the function. Removing dependence on a radio or GNSS signal can reduce the effect of jamming that signal. It does not remove the need for functioning sensors and navigation, and CSIS reporting shows vision-based systems can struggle over dense forest.[3][9]

Human control and the policy debate

The January 2023 edition of US DoD Directive 3000.09 requires appropriate levels of human judgment over the use of force and sets review procedures, with specified exceptions and a waiver route. That dated US policy is not a global ban or proof that all autonomous engagement is prohibited.[2]

The ICRC’s 2021 position calls for new binding rules, including prohibitions on unpredictable autonomous weapons and autonomous weapons targeting people. This is a humanitarian policy position advocating legal change. It must not be presented as a treaty already adopted by all states.[5]

The UN Convention on Certain Conventional Weapons (CCW) process scheduled expert-group sessions on lethal autonomous weapons for March and August–September 2026. The ICRC’s 2026 FAQ distinguishes autonomous navigation from autonomous targeting. Negotiating sessions, proposed rules and existing obligations are different things; none is a performance test of a named drone.[6][16]

▶Swift: autonomous racing research demonstrationWatch at the source · UZH Robotics and Perception Group ↗

The video opens only when you follow this link. No third-party video resources load here.[14]

Original schematic by DroneLedger editorial team · not to scale

Continue with our records

• Lancet (Izdeliye 51/52/53)
• Shahed-136 / Geran-2
• Switchblade 300 / 600
• IAI Harop
• ZALA KUB-BLA
• FPV kamikaze quadcopter
• Ukraine says new interceptor drones engage Shahed-type drones autonomously
• ZALA
• Shield AI