BRENNEN

Industries / Transport and vehicles / Aerospace and defense / Drones

Flight control and autonomy

Flight control is the onboard computer and software that keep a drone level and on course; autonomy is everything it does without a pilot's hands, from holding a hover to dodging trees and flying a whole route.

6

Milestones achieved, from 2009 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2009

    An open-source autopilot

    In May 2009 Jordi Muñoz and 3D Robotics released the first ArduPilot board, and that November Muñoz created the ArduPilot code repository. Anyone could read, change and reuse the code, and it grew to fly planes, multicopters, ground vehicles and underwater craft. A 2024 peer-reviewed paper describes ArduPilot as "installed on millions of small UAVs".

    ardupilot.orgfrontiersin.org

    Credit

    Jordi Muñoz; Chris Anderson, who started the DIY Drones community in 2007; the ArduPilot developers

  2. 2013

    An open autopilot board

    A student project at ETH Zurich, led by Lorenz Meier, produced the PX4 flight software, released with 3D Robotics in July 2012, and the Pixhawk autopilot board, released in November 2013, along with MAVLink, a common language for drones and ground stations to talk to each other. Pixhawk is now kept as a set of open hardware standards for autopilots, camera payloads and smart batteries under the Dronecode Foundation, with chip and drone makers such as NXP and Freefly taking part.

    pixhawk.orgardupilot.org

    Credit

    Lorenz Meier and the ETH Zurich team; 3D Robotics

  3. 2018

    A drone that dodges trees

    Skydio's R1, announced in February 2018 at about $2,500, followed a person on its own while steering around obstacles, using 12 of its 13 cameras to see all around it and an Nvidia TX1 graphics chip on board to process the images. The reviewer DC Rainmaker ran it along forest trails in Western Australia: it did not crash, but it clipped small branches, kept a 1 m clear space around itself and often stopped where a pilot could have passed. Skydio warned that small leafless branches might not be seen.

    LessonCamera-based avoidance worked among trees, and thin branches stayed its weak spot.

    spectrum.ieee.orgdcrainmaker.com

    Credit

    Skydio (co-founder Adam Bry, with Abe Bachrach and colleagues)

  4. 2018

    Jamming downs a light show

    On 27 October 2018, 46 of the 100 drones flying a choreographed light show over Victoria Harbour in Hong Kong fell into the water. The South China Morning Post attributed it to GPS jamming, interference that drowns out satellite positioning signals, and a Hong Kong Tourism Board official put the damage at no less than HK$1 million (US$127,500).

    LessonDrones that hold their place by satellite alone need a safe response when the signal is lost.

    scmp.com

    Credit

    South China Morning Post (reporting); Hong Kong Tourism Board (damage figure)

  5. 2023

    AI beats racing champions

    Swift, a racing drone flown by software trained through reinforcement learning (trial and error, mostly in simulation) and using only its onboard sensors, raced three human champions head to head: Alex Vanover, Thomas Bitmatta and Marvin Schaepper. It won 15 of 25 races and set the fastest recorded race time; of its 10 losses, 4 were collisions with the opponent, 4 with a gate and 2 were slower laps. The results were peer reviewed and published in Nature.

    nature.com

    Credit

    Elia Kaufmann, Davide Scaramuzza and colleagues at the University of Zurich's Robotics and Perception Group, with Intel Labs

  6. 2025

    AI wins an open race

    At the A2RL x DCL Autonomous Drone Championship in Abu Dhabi in April 2025, every team flew the same organizer-supplied drone, with an Nvidia Jetson Orin NX computer, one forward-facing camera and an inertial sensor. TU Delft's MAVLab won the autonomous race and then beat human pilots, which the organizers called the first win of its kind; TU Delft says it beat three former Drone Champions League world champions. Its neural network sends commands straight to the motors, a method first developed at the European Space Agency.

    LessonSwift's 2023 wins came on a lab track with its own hardware; this one came in an official race on equipment the organizers chose.

    aetoswire.commavlab.tudelft.nl

    Credit

    TU Delft MAVLab (Christophe De Wagter, Guido de Croon); organizers A2RL and DCL

Watch for

  1. A flying home camera

    Amazon says Ring Drone Cam, an indoor security camera that flies, maps a home from a walkthrough with software it calls AI Pilot, flies to preset points when a Ring sensor or alarm is set off, and steers around furniture on its own. It is listed at $249.99 for US shipping from 27 January 2027, and Amazon notes it does not yet have the FCC authorization it needs to be sold.

    Achieved whenit ships and an outside reviewer publishes how often it reaches its points without hitting anything (see the Note "A camera on patrol").

    aboutamazon.com

    Claimed by

    Amazon, November 2026

  2. AI reads crash logs

    A developer posting as furkanisikay released ardupilot-mcp, an open-source tool that lets an AI assistant such as Claude Desktop, or a local model, explain why an ArduPilot drone crashed. Sixteen fixed checks with set thresholds decide each finding, and the model explains it in plain words with a link to the ArduPilot manual. The author says it was validated against 40 real crash logs.

    Achieved whenexperienced log reviewers score its answers on crash logs it has not seen and publish the tally.

    discuss.ardupilot.org

    Claimed by

    furkanisikay on the ArduPilot forum, July 2026

Open

  1. 01

    An answer key for crashes

    A public set of at least 200 ArduPilot and PX4 crash logs whose cause experienced reviewers settled in public forum threads, each labeled with that cause and linked to the thread, then a blind score of AI diagnosis tools against it.

    The brief
    Credit · unclaimed
  2. 02

    Thin-obstacle tally

    A filmed test of how well drones sold with obstacle sensing stop for thin things: cords and cables of set widths, such as 1, 3 and 6 mm, and bare branches, hung on stands in a netted sports cage or an open private field with no one inside. Each drone approaches each obstacle at fixed speeds 20 times, and the stops, swerves and hits are tallied.

    The brief
    Credit · unclaimed

Open 01 · the brief

An answer key for crashes

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of at least 200 ArduPilot and PX4 crash logs whose cause experienced reviewers settled in public forum threads, each labeled with that cause and linked to the thread, then a blind score of AI diagnosis tools against it.

Why it is still open

ArduPilot made AI-assisted log diagnosis a 2026 Google Summer of Code project, and in February 2026 a maintainer told a student that no dataset of failed flight logs existed, only the forum. In March 2026 a diagnosis engine run blind on 34 forum logs was shown by reviewers to blame "compass interference" where the cause was thrust loss. The largest labeled set found, ArduCrash from a 2024 Frontiers paper, holds eight logs.

discuss.ardupilot.orgdiscuss.ardupilot.orgfrontiersin.org

Built with what exists

pymavlink and pyulog to read ArduPilot and PX4 logs; the ArduPilot forum's crash threads; existing tools such as Mission Planner's Auto Analysis, ardupilot-mcp and the student-built ALDA; general AI models such as Claude Fable 5.2 (see the Note "Three working devices"); an agent that cites the source behind each finding (see the Note "Findings that cite notes").

  1. First step

    Collect 30 forum threads from the past two years in which a developer named the root cause, download their logs and write each cause down in one line.

  2. Done looks like

    200 labeled logs published with a second reviewer's agreement rate on the labels, and at least three tools scored, with accuracy for each kind of failure.

Open 02 · the brief

Thin-obstacle tally

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A filmed test of how well drones sold with obstacle sensing stop for thin things: cords and cables of set widths, such as 1, 3 and 6 mm, and bare branches, hung on stands in a netted sports cage or an open private field with no one inside. Each drone approaches each obstacle at fixed speeds 20 times, and the stops, swerves and hits are tallied.

Why it is still open

Makers warn about it but publish no rates. Skydio's guide for its X2 says thin wires and branches are harder to detect in its Close and Minimal avoidance settings, and the 2018 R1 review above saw small branches missed. Researchers agree: Kolbeinsson and Mikolajczyk (September 2025) say wires pose "a unique and intricate problem" for drone vision because of their slender profile, and their paper builds a new wire detector; it does not test drones on sale. Reviewers such as DC Rainmaker have filmed side-by-side runs among trees (Skydio 2 against DJI Mavic 3, 2021), with no counts. A search found forum anecdotes but no controlled, tallied test of shipped drones against thin obstacles.

support.skydio.comdcrainmaker.comarxiv.orgdcrainmaker.com

Built with what exists

Three to five drones with obstacle sensing; a netted cage or open private field; cords and cables of known width on stands, well away from any power line; the drones' logs for speed; two cameras to film each approach from the side.

  1. First step

    Hang a 3 mm cord and a 10 mm rope between two stands in an open field and fly one drone toward each at 2 m/s ten times, filming from the side.

  2. Done looks like

    For each drone, obstacle and speed, the counts of stops, swerves and hits out of 20 and the closest distance measured from video, published with the clips.