Every DJI Mavic sold in recent years is marketed with the word "autonomous" somewhere in its feature list. Return-to-home, obstacle avoidance, waypoint routing — none of that constitutes autonomy in any rigorous engineering sense. The gap between marketing usage and technical reality is wider than most coverage suggests, and it matters more as UAS capability climbs.

Autonomy is not binary. It is a spectrum, and depending on which community is doing the measuring — academic roboticists, aviation regulators, or the Department of Defense — the spectrum gets carved differently.

Automation Is Not Autonomy

Beer, Fisk, and Rogers, writing in Frontiers in Psychology, define autonomy as "the extent to which a robot can sense the environment, plan based on that environment, and act upon that environment, with the intent of reaching some goal without external control." An automated drone follows pre-programmed instructions. An autonomous drone reasons about its surroundings in real time. Return-to-home executes a stored GPS coordinate. Waypoint navigation traverses operator-entered coordinates. These are automation — executing scripts, not interpreting situations.

"A drone that depends on GPS for stability is not autonomous. It is automated within a permissive environment. Remove the permission, and you remove the autonomy."

Exyn Technologies, whose systems operate in underground mining environments where GPS is unavailable and geometry is unmapped, draws the line between waypoint missions (requiring prior environmental knowledge and operator planning) and autonomous exploration via SLAM — Simultaneous Localization And Mapping — which builds real-time maps and determines routes independently in zero-light, GPS-denied space. The industry uses "autonomous" because, as one trade analysis noted, "the word sells better."

Five Frameworks, None Interoperable

The drone sector lacks a canonical standard equivalent to SAE J3016 for ground vehicles. Five distinct frameworks coexist, each built for a different problem domain.

Sheridan-Verplank (1978), the intellectual ancestor of all later work, was developed for underwater teleoperation and runs 10 levels from pure manual control to full autonomy. LORA (Levels of Robot Autonomy) adapts it with measurable metrics: neglect time — how long a system operates before human input is required — and subtask completion ratio, giving engineering teams something quantifiable to test against.

NIST ALFUS (Autonomy Levels for Unmanned Systems) deliberately resists single-axis numbering. Rather than assigning a level, ALFUS evaluates systems across three axes simultaneously: Mission Complexity, Environmental Difficulty, and Human Independence. A drone may operate at high human independence in a simple open-field survey and require effective teleoperation in a GPS-denied, obstacle-dense environment. The three-axis model captures that context-dependence; a numbered scale cannot. NIST now acknowledges the framework has not kept pace with field development.

The SAE-adapted aviation framework, developed by the European Cockpit Association from SAE J3016, demarcates at Level 4: Levels 0-3 require a Pilot-in-Command with flying skills; Levels 4-5 require only a Mission Commander. The ECA states that true autonomy — "where a system determines its own missions, makes independent decisions, and engages in strategic planning" — is "realistically not feasible in the near- to mid-term." Core principle: "A competent human must always be in command."

Exyn's aerial-specific framework adds sub-levels to a six-point scale: Level 4A covers independent exploration using onboard sensors; 4B adds reasoning about obstacle identity; 4C handles high-level obstacle classification. A system at Level 3 under SAE-adapted criteria might sit at Level 4 under Exyn's scale and at Level 8 in Sheridan-Verplank — the frameworks do not translate cleanly.

DoD's Human-Loop Taxonomy

DoD Directive 3000.09, originally issued in 2012 and updated January 2023, defines autonomy not by technological sophistication but by the human operator's role in target selection and engagement. A semi-autonomous weapon system ("human in the loop") engages only targets pre-selected by a human — a category that includes fire-and-forget guided missiles. A human-supervised autonomous weapon system ("human on the loop") allows an operator to monitor and halt engagement without requiring per-action approval. A lethal autonomous weapon system ("human out of the loop") selects and engages targets without further intervention after activation.

The directive requires that systems "will be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force." DoD has explained that "appropriate" is a flexible term that reflects the fact that there is not a fixed, one-size-fits-all level of human judgment that should be applied to every context — which Human Rights Watch noted leaves enforcement without clear teeth. The 2023 update removed references to "control," with a DoD official explaining the change. Reviews required before development and fielding can be waived in cases of urgent military need; as of 2019, no weapon system was publicly documented as having undergone the required senior-level review.

More than 70 countries as well as NGOs and the ICRC have called for a binding treaty with prohibitions and restrictions on lethal autonomous weapons under the U.N. Convention on Certain Conventional Weapons since 2014.

Where Human Control Actually Breaks

The MQ-9 Reaper, the U.S. military's frontline UAS, depends on satellite communications for ground control — sophisticated teleoperation, not autonomous operation, and those links are vulnerable in contested airspace. In Ukraine, drone lifespans have averaged nine days partly because ground control links are susceptible to jamming, making human-in-the-loop operation a tactical liability at modern combat tempo. As defense analyst Caitlin Lee notes, "The machine can move faster than that" — making human-in-the-loop operation a tactical liability at modern combat tempo.

At the high end of the spectrum, DoD's 2016 Perdix demonstration flew 103 micro-drones with "collective decision-making, formation flying and self-healing properties" via a distributed-brain architecture with no designated leader. Ukraine's Swarmer system coordinates 3 to 25 drones per mission across 100-plus documented combat deployments, reducing required operators from nine to three. China demonstrated coordination of up to 200 fixed-wing drones from mobile platforms with single-operator control in its Swarm I program. The Pentagon's Replicator program received $500 million in FY2024 to field thousands of inexpensive autonomous platforms at scale.

Commercial systems sit well below this threshold. DJI's automated waypoint flights and Skydio's computer-vision obstacle avoidance occupy the automation-to-conditional-autonomy range. Percepto and Airobotics drone-in-a-box platforms — scheduled missions with remote monitoring — represent the current commercial ceiling. True Level 5 autonomy, where a system self-governs against mission objectives without human involvement, is not deployed in any commercial application.

The honest question when evaluating any UAS autonomy claim is not "is it autonomous?" but three questions at once: autonomous at what task, under what environmental conditions, and who remains in the loop — and what happens to the mission when that loop is severed.

Sources