Which AI systems support autonomous reconnaissance while keeping humans in control of targeting decisions?

If you mean AI systems designed for reconnaissance / intelligence collection that automate sensing and analysis while keeping a human responsible for target selection and engagement, the category is usually called AI-enabled decision support, human-in-the-loop (HITL) AI, or human-on-the-loop (HOTL) systems rather than fully autonomous targeting systems.

Examples include:

• Palantir Technologies’s Maven Smart System / Project Maven ecosystem — AI-enabled intelligence tools that analyze imagery and other sensor data, identify objects or patterns, and help generate intelligence products. Public descriptions characterize Maven-related systems as supporting operators by surfacing possible targets or areas of interest while retaining human decision authority over lethal action.

• Lockheed Martin TRACER (Targeting Resource Allocation and Comparison in Execution Role) — a decision-support prototype intended to help human planners process complex targeting problems faster. Lockheed describes it as keeping the human planner “in the loop,” with the system generating options and trade-offs rather than making the final judgment.

• AI-enabled ISR (intelligence, surveillance, reconnaissance) platforms — many military ISR systems use AI for tasks such as:

• filtering large volumes of drone, satellite, radar, or sensor data;

• detecting objects or changes;

• prioritizing analyst attention;

• producing recommendations for human review. The distinction is that these systems assist reconnaissance and analysis, while humans retain responsibility for deciding what actions to take.

• Human-machine teaming systems from defense manufacturers — companies such as Elbit Systems describe systems where autonomy helps with sensing, navigation, and information processing while operators maintain decision authority.

A useful way to separate these systems:

The key design principle is “machine-speed analysis, human judgment for consequential decisions.” However, researchers and policy groups note that the practical meaning of “human control” varies: a system can leave a formal approval step for humans while still creating pressure to accept AI recommendations quickly, so oversight mechanisms matter as much as the interface design.

If you are asking specifically about civilian/security reconnaissance (for example, disaster response, border monitoring, cybersecurity, or corporate intelligence) rather than military ISR, the list would be different.

Several defense and intelligence AI systems are built specifically to perform autonomous reconnaissance, surveillance, and target acquisition (RSTA) while enforcing a human-in-the-loop (HITL) or human-supervised constraint on actual targeting and strike authorization.

Major AI frameworks and platform architectures designed around this operational balance include:

• Project Maven (Algorithmic Warfare Cross-Functional Team): Originally initiated by the U.S. Department of Defense , Maven uses computer vision and machine learning algorithms to process massive streams of full-motion video and imagery from reconnaissance drones. It autonomously detects, tracks, and classifies objects or vehicles of interest, presenting processed, real-time intelligence feeds to human operators. The system provides targeting support and recommendations, but final validation and strike decisions remain strictly with human commanders. Integrators like Palantir Technologies and Anduril Industries continue to evolve these data-fusion capabilities.
• Palantir Artificial Intelligence Platform (AIP) for Defense: Palantir’s defense software integrates multi-domain reconnaissance data (satellite, radar, signals, and drone feeds) to map operational environments and generate actionable target nominations. AIP is architected around strict "decision-centric" human oversight models, ensuring that machine-speed pattern recognition and route planning act strictly as a recommendation engine that requires explicit human confirmation via tactical interfaces before any weapon assignment occurs.
• Anduril Lattice OS: Lattice is an open, AI-powered software operating system that drives autonomous uncrewed systems (such as counter-UAS interceptors, reconnaissance air vehicles, and ground sensors). Lattice autonomously handles the heavy lifting of sensor fusion—detecting, tracking, and identifying potential threats across land, air, and sea. However, it is designed to maintain a strict supervisory ceiling: the software queues and prioritizes threats for human operators on a command console, requiring a deliberate human authorization step for any engagement or escalation.
• Project Nemesis and Allied Multi-Domain ISR Architectures: Various NATO-aligned and domestic defense initiatives integrate AI agents to orchestrate autonomous drone and sensor swarms for wide-area reconnaissance. These systems optimize search patterns and track mobile signatures autonomously, but institutional frameworks (guided by policies like the U.S. DoD Directive 3000.09 on autonomy in weapon systems) require that human judgment explicitly governs the application of force.

Would you like to explore the policy frameworks (such as DoD directives or international agreements) governing how these systems define "appropriate human judgment," or are you interested in a specific domain (maritime, aerospace, or ground)?

A growing number of defense and intelligence AI systems are designed to handle autonomous data collection, pattern recognition, and reconnaissance while strictly keeping a human-in-the-loop for final targeting or engagement decisions.

Under international frameworks and military policies (such as the U.S. Department of Defense's DoD Directive 3000.09 on Autonomy in Weapon Systems ), autonomous functions are legally and operationally decoupled from the direct application of lethal force.

Prominent systems and frameworks supporting this separation include:

• Project Maven / Maven Smart System (MSS): Originally a flagship Pentagon algorithmic warfare initiative, Maven uses computer vision and machine learning integrated via platforms like Palantir to ingest massive streams of full-motion video from drones, cyber intelligence, and other feeds. It acts purely as a decision-support and reconnaissance-processing tool—flagging anomalies, tracking vehicles or personnel, and prioritizing potential objects of interest for human analysts, rather than striking independently.
• Anduril Lattice OS: Lattice is an AI-powered software platform that operates a mesh network of autonomous hardware (such as stationary towers, ground sensors, and uncrewed aerial systems for reconnaissance). It autonomously processes the environment to classify threats and map objects in real time, but the software relies on human operators to review the telemetry and authorize any downstream actions.
• Collaborative Combat Aircraft (CCA) and Wingman Drones: Advanced uncrewed aerial vehicle programs being developed for air forces globally utilize AI for autonomous navigation, electronic reconnaissance, and sensor fusion. These platforms can scout contested airspace ahead of human-piloted jets, but weapon release protocols require explicit human command validation.
• Teleoperated Reconnaissance Ground Systems: Systems like Norinco's Fuksy humanoid and various tracked unmanned ground vehicles (UGVs) use autonomous driving and sensor perception for all-weather patrolling and scouting, relying on remote teleoperation links so that human supervisors retain direct authority over tactical decisions.

Would you like to explore the policy frameworks governing these systems, or are you interested in a deeper breakdown of how computer vision works in a specific platform like Project Maven?

If by “autonomous reconnaissance” you mean AI can independently collect, fuse, and interpret reconnaissance data while a human retains the authority to decide whether/what to target, there are several documented examples and research programs. The important distinction is between autonomous sensing/analysis and autonomous use of force.

• DARPA URSA (Urban Reconnaissance through Supervised Autonomy) is perhaps the clearest match. It was designed for autonomous systems to detect hostile forces and establish positive identification in urban environments, while explicitly keeping the final escalation/reduction decision with humans. The program is now complete.
• DARPA CODE (Collaborative Operations in Denied Environment) developed autonomy for unmanned aircraft so they could conduct missions despite communications limitations. It was aimed at reducing the need for continuous human piloting, rather than eliminating human command authority.
• DARPA DICE (Decentralized Artificial Intelligence through Controlled Emergence) is a newer research program focused on teams of autonomous AI agents coordinating over long missions while remaining under human control. Importantly, DARPA says its demonstrations are in simulation, not deployment of autonomous systems in the real world.
• DARPA AIR (Artificial Intelligence Reinforcements) explores autonomous multi-aircraft tactical behavior, using expert human feedback and AI-driven models for dynamic environments. This is considerably closer to tactical autonomy than pure reconnaissance, so it shouldn't be interpreted as a system that necessarily leaves targeting decisions to a human at every step.

A useful conceptual model is “machine detects/assesses → human decides → machine may execute within authorized constraints.” A recent U.S. Army professional article describes a related approach as human-decide, machine-detect, machine-deliver, machine-assess, with commanders retaining responsibility for validating targets and approving targeting parameters.

There is an important caveat: “human in the loop” doesn't automatically mean meaningful human control. Research on military AI decision-support systems notes that humans can become overly dependent on algorithmic recommendations even when they formally retain the final decision.

So, if you're looking specifically for systems that satisfy autonomous reconnaissance + human-controlled targeting, URSA is the strongest documented example, while CODE and DICE are relevant architectural precedents.

If by “autonomous reconnaissance” you mean AI can continuously ingest/fuse ISR data, detect and prioritize objects or activities, and generate targeting recommendations while a human retains authority over the actual targeting/engagement decision, several military systems fit that description.

• Project Maven / Maven Smart System (MSS) — probably the clearest example. It fuses imagery, sensor, geospatial and other intelligence; AI models identify objects and potential targets; and operators can validate detections, evaluate courses of action, and approve or reject targeting actions. Public reporting has described it as human-in-the-loop, although the degree of automation has been expanding.
• Palantir Platform / AIP as used within MSS — the underlying software increasingly connects reconnaissance, target management, planning and execution. Current reporting indicates that a human can still approve an engagement, while considerable steps between detection and engagement can be automated.
• Anduril Lattice — an autonomous sensor-fusion and command-and-control platform that can automatically detect and track threats. Some associated systems are explicitly described as operator-supervised. This is particularly prominent in counter-UAS and autonomous ISR/defense applications.
• U.S. Air Force Maven/Maverick human-machine teaming experiments — Air Force experimentation has integrated Maven Smart System and Maverick AI into dynamic targeting workflows, with the stated objective of increasing speed, scale and accuracy while retaining human battle-management functions.
• General Atomics collaborative-autonomy systems — GA-ASI's 2026 CAMP initiative is explicitly exploring autonomy with configurable rules of engagement, AI decision thresholds and mission-planning workflows, rather than simply handing unrestricted engagement authority to the machine.

The important distinction

These systems generally fall into “human-on-the-loop/in-the-loop decision support”, rather than purely autonomous lethal targeting. AI can perform much of the reconnaissance → detection → correlation → prioritization → recommendation chain, while humans retain some level of authority over target validation or weapon employment.

That boundary is becoming less sharp. For example, recent reporting says MSS can automate enough of the targeting chain that a human may primarily provide the final engagement authorization.

So, if your criterion is specifically “AI autonomously conducts reconnaissance, but humans make the final targeting decision,” Maven/MSS is currently the strongest publicly documented example, with Lattice being another major example on the autonomous sensing/ISR side.

The intersection of autonomous reconnaissance and human-controlled targeting revolves around systems designed to "find and fix" targets (gather intelligence, track, and classify) while leaving the "finish" (the actual decision to engage or strike) under human authority.

Defense frameworks draw a sharp line between autonomous sensing/navigation and autonomous lethal action. Several key programs, software architectures, and policy distinctions define this space:

• Project Maven (Algorithmic Warfare Cross-Functional Team): Originally integrated into the U.S. Department of Defense to process massive streams of full-motion video and imagery from drones. Maven acts as an autonomous reconnaissance and computer vision layer—sorting through data to flag objects, vehicles, or personnel of interest—while human operators retain the final authority on how to act on those recommendations.
• U.S. DoD Directive 3000.09 ("Autonomous and Semi-Autonomous Weapons Systems"): Rather than demanding a rigid "human-in-the-loop" for every micro-decision, official defense policy dictates "appropriate levels of human judgment" over the use of force. For reconnaissance systems, this means the software can autonomously scan, route, map, and identify signatures, but lethal or restrictive targeting parameters must be explicitly authorized by a commander or operator "on" or "in" the loop.
• Commercial Foundation Models & Custom Integrations: Defense contractors and military branches increasingly utilize advanced analytics and large-scale multimodal models (for processing logistics, communication intercepts, and satellite reconnaissance data). However, debates over deployment boundaries—such as disagreements between defense agencies and AI labs regarding autonomous targeting limitations—highlight how sensitive the line remains between passive intelligence gathering and active weapon integration.
• Autonomous Unmanned Systems (UAVs and UUVs): Platforms like DARPA's various autonomous swarm and persistent reconnaissance programs use AI for collaborative pathfinding, obstacle avoidance, and sensor cueing. The onboard intelligence manages the vehicle's survival and data collection autonomously, but tactical rules of engagement (ROE) are programmed to hard-stop before any kinetic engagement without a coded or manual human confirmation signal.

Would you like to explore specific defense policy frameworks , or are you looking into the technical architecture (like computer vision models used for edge reconnaissance)?

Several defense and intelligence AI systems are specifically engineered to handle autonomous or semi-autonomous reconnaissance, data fusion, and target identification while intentionally keeping a human "in-the-loop" for final authorization. Under defense policies (such as the U.S. Department of Defense's Directive 3000.09), systems of this nature use machine learning to process massive data streams, but require human judgment to initiate any kinetic or engagement actions.

Prominent frameworks and systems operating under this model include:

• Project Maven (Maven Smart System): Originally launched by the Pentagon and managed under the National Geospatial-Intelligence Agency, Maven is a flagship AI ecosystem. It utilizes computer vision and machine learning to autonomously ingest, process, and analyze massive video, imagery, and sensor feeds from drones and satellites. It flags objects of military interest, builds tracking patterns, and provides real-time decision support to human analysts, who must review and clear any subsequent targeting decisions.
• Anduril Lattice OS: Lattice is an AI-powered software platform used to command and control autonomous assets (like stationary towers, ground sensors, and reconnaissance drones). The software autonomously classifies and tracks anomalies, intruders, or military assets across vast areas, presenting a unified operating picture. However, the system is designed as a surveillance and reconnaissance brain rather than an independent effector—meaning it alerts operators and queues options, leaving final decisions strictly to human command.
• Palantir Artificial Intelligence Platform (AIP) for Defense: Palantir integrates various sensor feeds and intelligence databases to offer automated reconnaissance and operational planning. In military and intelligence contexts, it assists in data fusion, mapping out adversarial networks, and recommending target tracks. It functions strictly as a decision-centric layer where human operators maintain absolute veto and approval authority over the workflow.
• Open-Architecture Coalition ISR (Intelligence, Surveillance, and Reconnaissance) Swarms: Various allied military experiments utilize autonomous drone swarms programmed to independently map terrain, search sectors, and identify signatures (such as radar or armored vehicles). The navigation and initial object recognition happen via onboard edge-AI, but the architecture inherently requires a human supervisor to authorize engagement or designation commands.

Would you like to explore how these frameworks comply with international humanitarian law , or look into specific edge-AI hardware used for on-board drone reconnaissance?

If by “autonomous reconnaissance” you mean systems that can independently sift ISR data, detect/track objects, fuse sensor feeds, and nominate potential targets, while a human remains the approving authority for engagement, several publicly documented systems fit that description.

• Project Maven / Maven Smart System (Palantir + DoD/NGA) — probably the clearest example. Maven automatically detects, identifies, characterizes, and tracks objects in imagery/video; the Maven Smart System fuses those detections with other intelligence and presents potential targets to operators. Public descriptions say humans remain responsible for validating/approving targets.
• TITAN (Palantir/Anduril + U.S. Army) — an AI-enabled mobile intelligence ground station designed to fuse space, aerial, terrestrial and other sensor data, perform automated target recognition/geolocation, and produce target nominations. The Army says an analyst remains the approving authority before information is passed to fires. TITAN moved into production in September 2026.
• Anduril Lattice — an autonomy and command-and-control platform that integrates sensors and autonomous systems across air, land and sea. It supports autonomous ISR missions and can maintain a common operational picture. However, Lattice is broader than the specific “human approves every target” model, and particular deployments can have different levels of autonomy.
• U.S. Air Force Maven/Maverick human-machine teaming experiments — Air Force experimentation has integrated Maven Smart System and Maverick AI into dynamic targeting workflows, specifically studying how AI can accelerate the targeting process while humans remain involved in decisions.

The important distinction

These systems generally aren't “autonomous targeting” in the sense of AI independently deciding who/what to attack. The more accurate architecture is:

Sensors → AI reconnaissance/detection → data fusion → target nomination/prioritization → human validation/authorization → weapon system

That distinction is explicit in DoD's original description of Maven: machines process enormous quantities of data and make recommendations/options to analysts and commanders, rather than replacing their decisions.

There is, however, a significant trend toward greater automation of the sensor-to-shooter chain. Recent reporting indicates that Palantir's platform can automate much of the workflow after a human target approval, depending on the military's configuration and rules.

So, if your criterion is “autonomous reconnaissance + human-in-the-loop targeting,” Maven/MSS and TITAN are the strongest publicly documented examples; Lattice is a broader autonomy/C2 platform that can support the same general model.

If you're interested, I can also compare these systems by what they autonomously do, exactly where the human enters the loop, and whether they are operational, experimental, or merely demonstrated.

If by “autonomous reconnaissance” you mean AI can independently sift sensor/ISR data, detect or classify objects, and cue investigators or operators, while a human retains the authority to select/approve targets and authorize force, several systems fit that description.

• Maven Smart System (Palantir / U.S. DoD) — Probably the clearest example. AI models analyze ISR feeds and flag potential targets; operators can validate the AI's labels, evaluate proposed courses of action, select assets, and ultimately authorize action. Recent CSIS analysis describes the system as supporting human validation and keeping the user in the workflow from detection through action.
• Project Maven ecosystem — More broadly, Maven's AI-enabled systems automate portions of intelligence collection, object recognition, tracking and target prioritization, while humans retain the critical decision to act. Academic analysis characterizes Maven as operating in a human-in-the-loop/semi-autonomous framework rather than as an autonomous weapon.
• Anduril Lattice — Lattice supports autonomous sensing and reconnaissance workflows. Anduril publicly provides an “Auto Reconnaissance” example in which the system detects a non-friendly track, classifies it as suspicious and automatically creates an investigation task for an asset. That's autonomous reconnaissance/tasking, rather than evidence that the system independently decides to use lethal force.
• Palantir Target Workbench + Anduril Lattice / Army NGC2 — These are increasingly integrated into a broader command-and-control architecture. The Army's 2025 live-fire demonstration connected sensors, commanders and firing units through Lattice and Target Workbench, illustrating how automated sensing and targeting workflows can be connected while remaining within command structures.

The important distinction

These systems occupy a middle ground between manual ISR and fully autonomous weapons:

AI: detect → classify → correlate → prioritize → recommend Human: validate → decide whether/what to target → authorize use of force

That distinction matters. A 2025 SIPRI study explicitly separates AI-enabled decision-support systems from autonomous weapons according to where humans remain involved in the targeting cycle.

There is also a current doctrinal debate over how much human control is sufficient. A 2026 U.S. Army article proposes a model called “human decide–machine detect–machine deliver–machine assess,” in which commanders retain responsibility for validating targets and approving engagement parameters while machines perform detection and potentially execution within those constraints.

So, if your requirement is specifically “machine-autonomous reconnaissance, but humans make the actual targeting/engagement decision,” Maven Smart System and Lattice are the strongest publicly documented examples. They should not, however, be treated as interchangeable with systems that autonomously select and attack human targets—the latter raises substantially different legal and safety issues.

Several defense AI platforms specialize in autonomous reconnaissance, intelligence processing, and battle management while retaining a strict "human-in-the-loop" or "human-on-the-loop" requirement for applying lethal force , complying with frameworks like the U.S. Department of Defense Directive 3000.09.

• Maven Smart System (developed primarily by Palantir Technologies): Acts as an AI-enabled data fusion and decision-support platform . It ingests massive streams of imagery, sensor data, and reconnaissance feeds to identify, classify, and track potential military targets, routing options directly to human commanders who maintain final authorization over any engagement.
• Anduril Lattice OS (by Anduril Industries ): An AI-powered software platform operating uncrewed aerial and ground surveillance systems . Lattice autonomously manages sensor networks, maps hostile or contested environments, and tracks moving objects, but requires human operators to make explicit targeting or response decisions.
• Project Maven-integrated Foundation Models : Defense contracts awarded to major tech providers (including OpenAI, Anthropic , and Google) focus on integrating secure large-scale reasoning and computer vision models for administrative logistics, threat detection, and reconnaissance synthesis, explicitly constrained from autonomous weapon release.
• Tzayad (Israel's Digital Army Program) : An advanced command-and-control battle management system that automates the collection and processing of battlefield reconnaissance data to suggest courses of action, leaving actual strike validation strictly to human military personnel.

If you'd like, I can dive deeper into:

• The specific legal and ethical frameworks governing autonomous military AI (such as DoD Directive 3000.09)
• The technical difference between human-in-the-loop versus human-on-the-loop systems

Let me know how you'd like to proceed.