BRENNEN

Industries / Transport and vehicles / Aerospace and defense / Drones

Cameras and sensors

This is the work of the payload: the film cameras, heat cameras and laser scanners a drone carries, and the gimbals, motorized mounts that hold them steady.

5

Milestones achieved, from 1995 to 2021.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1995

    Film cameras on model helicopters

    At its Scientific and Technical Awards on 4 March 1995, the Academy of Motion Picture Arts and Sciences honored Emmanuel Previnaire of Flying-Cam for his concept and for mounting a motion-picture camera on a remote-controlled miniature helicopter. Film crews could get low, moving aerial shots without a full-size helicopter.

    journal.smpte.org

    Credit

    Emmanuel Previnaire, Flying-Cam

  2. 2013

    Camera drones come ready-built

    On 7 January 2013 DJI, a Chinese drone maker, released the Phantom, a $679 four-rotor drone that came assembled, held its position by satellite navigation and had a mount for a GoPro action camera. DJI called it the industry's first "Ready-To-Fly" satellite-navigated quadcopter and said it cost half as much as rival kits that had to be built and soldered.

    dji.com

    Credit

    DJI

  3. 2014

    Film sets get exemptions

    On 25 September 2014 the US Federal Aviation Administration (FAA) exempted six aerial filming companies from rules that had kept them from flying drones for pay. The Motion Picture Association of America put the requests forward. Pilots had to hold a private pilot certificate, keep the drone in sight, fly only over a closed set, inspect the aircraft before each flight and not fly at night. AOPA, the US pilots' association, reported that commercial drone approvals had until then been "extremely rare."

    aopa.orgdronelife.com

    Credit

    Astraeus Aerial, Aerial MOB, HeliVideo Productions, Pictorvision, Vortex Aerial and Snaproll Media; FAA

  4. 2021

    Thermal cameras pass lab test

    A thermal camera turns heat into a picture. DJI and FLIR announced the Zenmuse XT, a 640 by 512 pixel thermal camera on a DJI gimbal, on 10 December 2015, for firefighting, farming and inspection. In 2021 George Leblanc, Margaret Kalacska, J. Pablo Arroyo-Mora and colleagues tested three DJI thermal cameras (the Zenmuse XT-R, the Zenmuse XT2 and the Mavic 2 Enterprise Advanced) against a blackbody, a reference surface held at a known temperature. Each responded in a straight line from 5 to 40 °C, with a bias under 2 °C, inside the stated accuracy.

    produccioncientifica.usal.esdji.com

    Credit

    DJI and FLIR for the cameras; Leblanc and colleagues for the test

  5. 2021

    A one-take indoor flight

    "Right Up Our Alley" is an 87-second film shot in one continuous take by a cinewhoop, a small FPV drone with guarded propellers, flown through the Bryant Lake Bowl in Minneapolis. FPV means "first-person view": the pilot flies through goggles that show the drone's camera feed. CineD reported about 1.6 million YouTube views within a few days of its March 2021 release; the team needed a few attempts, and the sound was added afterward because propeller noise drowned it.

    cined.com

    Credit

    pilot Jay Christensen, director Anthony Jaska

Watch for

  1. Lidar ranges to 950 meters

    Lidar measures distance by timing laser pulses. DJI says its Zenmuse L3 lidar, released on 4 November 2025, detects objects of 10% reflectivity at up to 950 m (at a 100 kHz pulse rate in 100 klx daylight), gives vertical accuracy of 3 cm at 120 m and 5 cm at 300 m, and covers up to 10 km² in one flight at 300 m on its Matrice 400 drone.

    Achieved whena university, survey agency or transport department publishes a test of the L3 against surveyed checkpoints that meets these figures.

    enterprise.dji.com

    Claimed by

    DJI, November 2025

Open

  1. 01

    Blackbody tests for drone cameras

    A public bench test of the thermal cameras now sold for inspection drones, such as DJI's Mavic 3T, Matrice 30T, Matrice 4T and Zenmuse H30T and their Autel and Teledyne FLIR rivals. Each camera is pointed at a blackbody at set temperatures, distances and humidity, and every raw image and the error at each point are published.

    The brief
    Credit · unclaimed
  2. 02

    Shot plans against camera operators

    A blind test in which an AI agent turns written shot requests, such as "a slow reveal of the barn from behind the trees", into drone flight paths in a simulator, and professional aerial camera operators fly the same requests in the same simulated scene. Viewers rate the clips without knowing which is which.

    The brief
    Credit · unclaimed

Open 01 · the brief

Blackbody tests for drone cameras

Credit · unclaimed

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

What to build

A public bench test of the thermal cameras now sold for inspection drones, such as DJI's Mavic 3T, Matrice 30T, Matrice 4T and Zenmuse H30T and their Autel and Teledyne FLIR rivals. Each camera is pointed at a blackbody at set temperatures, distances and humidity, and every raw image and the error at each point are published.

Why it is still open

The only outside blackbody test found covers three older DJI cameras (Leblanc and colleagues, 2021). For current models the published figure is a rating, such as ±2 °C or ±2% for the Zenmuse H20T, measured "at 25 °C in a windless laboratory, 5 m away from the black body", as a UK retailer's knowledge base gives it. STR Drones publishes image sets from current DJI thermal cameras on Hugging Face, but they test how software reads the files' metadata, not temperatures against a reference. A GitHub search for blackbody tests of drone cameras returned nothing.

produccioncientifica.usal.esheliguy.comstrdrones-readme.static.hf.space

Built with what exists

A blackbody calibration source and a reference contact thermometer; the cameras fixed on a bench with propellers removed, so nothing flies; DJI's Thermal SDK and open readers for radiometric JPEG files; Python for the analysis.

  1. First step

    Rent one blackbody and borrow one drone thermal camera, measure five set points at 5 m indoors, and compare the method and results with Leblanc and colleagues' 2021 tests.

  2. Done looks like

    At least six current camera models, each measured at 10 set temperatures from 0 to 150 °C, at 2, 5 and 10 m, with five repeats per point, and a published table of bias and spread per model beside all raw files.

Open 02 · the brief

Shot plans against camera operators

Credit · unclaimed

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

What to build

A blind test in which an AI agent turns written shot requests, such as "a slow reveal of the barn from behind the trees", into drone flight paths in a simulator, and professional aerial camera operators fly the same requests in the same simulated scene. Viewers rate the clips without knowing which is which.

Why it is still open

ACDC (Lin and colleagues, September 2025) turns natural-language requests into drone paths in simulated indoor scenes and hardware-in-the-loop tests, but its abstract reports no comparison with human pilots and no viewer ratings. Bonatti and colleagues at Carnegie Mellon (2019) had ten viewers rank clips from their learned shot-choosing policies against fixed and random ones, not against human operators. CinePilot, a 2026 GitHub project, suggests the next shot from a live drone feed using Google's Gemini model, while the pilot still flies every shot.

arxiv.orgarxiv.orggithub.com

Built with what exists

PX4's software-in-the-loop simulator with Gazebo; 3D scenes built from photos with photogrammetry or Gaussian splatting; a vision-language model such as Claude Fable 5.2 (see the Note "Three working devices") or Gemini; ACDC's method as the baseline.

  1. First step

    Write 10 shot requests for one free simulator scene, have one camera operator fly them, and have an off-the-shelf model write paths for the same 10.

  2. Done looks like

    30 shot requests, each flown by the agent and by three human operators in the same scene, rated blind by at least 20 viewers, with all clips, paths and ratings published and the share of requests where the agent's clip is rated as good or better.