Investigating accidents across the uncrewed continuum

The pilot survives every drone accident, but the evidence scatters into software, ground stations, and RF spectrum. How autonomy and proliferation are rewriting air safety investigation.

At the 2025 annual seminar of the International Society of Air Safety Investigators, Concept Solutions program manager Jay Graser, a retired Air Force aircrew member with more than 2,500 flying hours, presented a paper with an unglamorous premise and far-reaching consequences: the air safety community is not yet trained for the accidents that uncrewed aircraft will produce. This article draws on that paper.

Thirteen to one

The FAA estimates that close to 2.8 million small drones were operating in the National Airspace System as of 2024: roughly 1.8 million recreational and just under a million commercial. Set that against approximately 215,000 civil manned aircraft and the ratio lands at thirteen to one. The imbalance is accelerating. Operators worldwide reported 19.5 million takeoffs in 2024, a 25 percent jump over the prior year, and the global drone market is projected to grow from $40.6 billion to $57.8 billion by 2030.

The safety data is following the fleet. Pilots filed 2,597 drone sighting reports in 2021, more than double the count from 2015, the first full year of tracking. In a single quarter, January through March 2025, another 410 sightings reached the FAA. Europe shows the same trajectory: the Eurocontrol ACUTE project recorded 3,833 encounters between crewed aircraft and drones over European cities in 2024, up from 1,743 a year earlier.

What makes these numbers hard to act on, Graser argues, is that "drone" is not a category. It is a continuum. The same word covers a sub-half-pound toy that cannot lacerate skin, a commercial aircraft over 1,000 pounds, a military platform over 30,000 pounds, and, on the horizon, remotely piloted cargo aircraft at the scale of a Boeing 747-400F. The FAA slices this continuum into Groups by size and operating altitude and into Categories by weight and injury potential; EASA labels civil drones C0 through C6. An investigator arriving at an occurrence involving "a drone" therefore knows almost nothing until the machine's place on the continuum is established. Everything downstream, from evidence sources to probable cause, depends on it.

The oldest assumption in the wreckage

Accident investigation grew up around a stable fact: the aircraft and its pilot occupy the same place. Uncrewed systems break that assumption, and the break cuts both ways. The upside is that no matter how badly the aircraft is destroyed, the pilot survives to be interviewed. The downside is that the century-old organizing principle of collision avoidance, see and avoid, no longer describes how the system actually works.

The disruption is twofold. The remote pilot has no cockpit perspective from which to execute an avoidance maneuver. And most drones are dramatically smaller than even the smallest manned aircraft, which means that by the time a manned crew visually acquires one, the window to maneuver may already have closed. The practical burden of separation shifts to the UAS operator, who is far more likely to detect the manned aircraft first than to be detected.

Beyond Visual Line of Sight (BVLOS) operations stretch the assumption further. Under BVLOS, the pilot and ground station may be far removed from the occurrence site. The investigator's first questions change accordingly: not "where was the pilot" but "where was the ground station, what did it record, and was the aircraft transmitting?"

Autonomy moves the root cause into code

Software has already forced its way into manned-aircraft investigation; MCAS on the Boeing 737-8 made that unmistakable. Uncrewed systems compress the trend. Even a simple recreational drone flies on thousands of lines of code, and the question of who, or what, was flying at the moment of the accident now has six possible answers. JARUS, the Joint Authorities for Rulemaking of Unmanned Systems, defines automation levels running from Level 0, where the human fully controls the aircraft, to Level 5, full automation, where the human cannot intervene in real time at all.

Where an occurrence falls on that scale reshapes the investigation. At Levels 0 through 2, the familiar human-factors toolkit applies: training, workload, situational awareness, fatigue, over-reliance on assistance functions. At Levels 3 through 5, the inquiry moves into the machine itself: the algorithms, the sensor data feeding them, and the design decisions behind both. A lost-link event during BVLOS flight is the canonical example. If the aircraft was executing a pre-programmed recovery, autonomously returning to its origin or diverting to a designated safe landing point, the focus belongs less on the pilot and more on the programming.

Two structural differences from manned aviation sharpen the software question. Aviation has relied on automation since Sperry invented the autopilot in 1912, but there was always a human aboard with a vested interest in the outcome. Self-preservation as a design assumption disappears when the operator stands on the ground. The timing problem is just as fundamental: an automated see-and-avoid function routed through the ground station inherits the control link's latency, so collision avoidance that works in the split seconds available has to run at the edge, onboard the aircraft, drawing on cameras, LiDAR, radar, and infrared sensors that extend "sight" well beyond human vision.

The hardware compounds the challenge. To save cost and weight, drones frequently fly hardware and software that were never flight-certified, airframes built from lightweight composites that can fail abruptly outside design limits and are difficult to detect with conventional radar, and lithium polymer batteries that carry an elevated fire risk. Each is a failure mode that manned-aviation investigators rarely confront in this combination.

The flight recorder is now a distributed system

There is no orange box to recover. The data an investigator needs, Graser points out, is distributed across the system: video and telemetry in onboard memory, command histories and link-quality logs at the ground station, and records held by third parties such as law enforcement agencies that may have been tracking drones in the area. The discipline is the same one applied to flight recorders in a manned accident. Identify every data source early and lock it down before it is overwritten, powered off, or walked away.

Detection infrastructure adds three more evidence streams, none of which depends on the drone cooperating. Radar tracks the aircraft itself, even when optimized for other missions. Optical sensors, especially cameras cued by radar, provide imagery in conditions of limited visibility. Radio frequency sensors intercept the signals between ground station and aircraft, and through triangulation can locate the drone and, in some cases, the operator as well.

Eyewitness accounts deserve more caution here than usual. Law enforcement officers report that planets, crewed aircraft, and even low-earth-orbit satellites have been confidently identified as drones. With drones constantly in the media, witnesses are primed to see one; that psychology is itself a variable the investigator has to weigh.

Regulation is accelerating toward the investigator

The regulatory environment is moving fast, and every expansion of operations expands the investigative problem space before any casebook exists. The Unleashing American Drone Dominance Executive Order, signed in June 2025, directs the FAA to prioritize BVLOS rulemaking, streamlines approvals for advanced drone technologies, and emphasizes investment in autonomy, artificial intelligence, and secure communications. The forthcoming FAA Part 108 framework will replace today's case-by-case waivers with standardized Operational Permits and Operational Certificates, opening the door to complex missions such as large-scale delivery.

The waiver curve shows the demand that Part 108 will formalize. In 2024 the FAA granted 809 waivers to Part 107, a 256 percent increase over 2023, and a quarter of them were for BVLOS operations. At the far end of the continuum, NASA in 2024 began a testing campaign for remotely piloted cargo aircraft, including a successful demonstration of ground-based radar ensuring safe separation during taxi, takeoff, and landing. Full-size remotely piloted aircraft are already flying test runs to villages in Alaska, exactly the kind of remote, low-density environment where they are expected to operate first.

Staying ahead of the continuum

The number of potentially reportable incidents will rise as uncrewed aircraft proliferate, and the community of trained investigators has to keep pace. Keeping pace means learning the technology well enough to ask the right questions across all four axes of the continuum: small to large, recreational to commercial, low automation to high automation, low risk to high risk. The groundwork exists. The ISASI Unmanned Aircraft System Handbook and Accident/Incident Investigation Guidelines, published in 2015, remains relevant a decade on.

Investigators can also shape the rules they will one day investigate under. FAA Notices of Proposed Rulemaking typically carry a 45 to 60-day comment window, and ICAO proposals a 90-day period. The professionals who understand what an uncrewed accident investigation actually requires are precisely the people who should be commenting. The drones, as Graser concludes, are here to stay. The question is whether the investigative discipline arrives ahead of the accident, or after it.

This article is based on "Investigating Uncrewed Vehicles: The Continuum and What to Consider," presented by Jay F. Graser at the 2025 ISASI Annual Seminar. Jay is an Information Systems Program Manager at Concept Solutions and a retired U.S. Air Force aircrew member with over 2,500 flying hours; he manages the FAA Command and Control Communications (C3) Support contract, providing situational awareness for the FAA Operations Centers. Contact our team for a copy of the full paper.