💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

원문 렌더가 준비되기 전까지 텍스트 가이드로 표시합니다.

In spring 2026, drones have fully moved from being "aerial cameras" to being "aerial computers." The US has effectively legislated a ban on federal procurement of DJI components, and as the Ukraine war enters its fifth year, the share of military autonomous drones has exploded. Agriculture, forestry, logistics, and maritime rescue are all shifting from "one operator per drone" to "one operator running ten autonomous drones." At the same time, developers face more complicated choices than ever: PX4 vs ArduPilot, Pixhawk vs ModalAI, and what to use after AirSim was sunset.

This article maps the entire drone development stack as of May 2026. We split it into four buckets — open autopilots, commercial SDKs, autonomy middleware, simulation — and walk through each with selection criteria, including the Korean and Japanese local ecosystems.

1. The 2026 Drone Development Map — Open Autopilot / Commercial SDK / Autonomy / Simulation

You cannot reduce drone development to a single stack because the layers are radically different. "Doing drone development" can mean porting PX4 firmware, controlling a camera with the DJI Mobile SDK, or driving an autonomous inspection through the Skydio Autonomy API. In 2026 the cleanest split is along four axes.

- Open autopilots — PX4 (Dronecode / Linux Foundation), ArduPilot (GPLv3, originally BSD-spirited)

- Commercial SDKs — DJI Mobile/Onboard/Payload SDK, Skydio Autonomy SDK, Parrot Olympe, Autel SDK

- Autonomy middleware — ROS 2 + PX4, Skydio Autonomy, Auterion Suite, Anduril Lattice

- Simulation — Microsoft AirSim (sunset 2023), NVIDIA Isaac Drone Sim, Gazebo Sim, jMAVSim, Flightmare

| --- | --- | --- | --- | --- |

The rest of the article walks each axis with its flagship stack.

2. PX4 Autopilot — Open Standard Under Dronecode (Linux Foundation)

PX4 started at ETH Zurich's Computer Vision and Geometry Lab and has been governed since 2014 by the Dronecode Foundation, now a Linux Foundation project. In 2026 it is the most-adopted open autopilot, and because it is BSD-3-Clause, you can embed it in commercial products without forcing source disclosure.

The PX4 architecture stacks a NuttX RTOS (or Linux with PREEMPT_RT) under a uORB publish/subscribe message bus, and every module (sensor driver, EKF2 attitude estimator, attitude/position controller, MAVLink router) talks through uORB topics. PX4 effectively behaves like a miniature ROS internally.

// PX4 — module that subscribes to the vehicle_attitude uORB topic

#include <uORB/uORB.h>

#include <uORB/topics/vehicle_attitude.h>

int my_module_main(int argc, char *argv[]) {

int fd = orb_subscribe(ORB_ID(vehicle_attitude));

struct vehicle_attitude_s att;

while (true) {

if (orb_check(fd, &(bool){false}) == 0) {

orb_copy(ORB_ID(vehicle_attitude), fd, &att);

PX4_INFO("q=[%.3f %.3f %.3f %.3f]", (double)att.q[0],

(double)att.q[1], (double)att.q[2], (double)att.q[3]);

}

return 0;

}

As of May 2026 the stable line is PX4 v1.16, and the headline changes are: (1) ROS 2 integration is standardized through the uXRCE-DDS bridge; (2) the PX4-Autopilot repo now ships first-class support for ARK Electronics' ARKV6X FMU board; (3) a safety-certified track (DO-178, DO-254) is split out as PX4 Pro and driven mainly by Auterion.

The main weakness of PX4 is that for some niche scenarios — indoor GPS-denied flight, agricultural spray patterns — ArduPilot has more ready-made modules.

3. ArduPilot — The Oldest Open Autopilot, GPLv3

ArduPilot started in 2007 in the DIY Drones community, which makes it the oldest open-source autopilot in 2026, eighteen years in. The biggest difference from PX4 is the license: ArduPilot is GPLv3, which means if you ship a drone with modified ArduPilot firmware, you have to publish your changes. That's why PX4 dominates in commercial products while ArduPilot dominates in research, hobby, and government.

ArduPilot builds ArduCopter (multirotor), ArduPlane (fixed wing), ArduRover (ground vehicle), ArduSub (underwater), AntennaTracker, and Blimp from a single code base. PX4 is aircraft-centric; ArduPilot is for "anything that moves."

// ArduPilot — adding a custom Copter mode (Copter/mode_custom.cpp)

#include "Copter.h"

#include "mode.h"

bool ModeCustom::init(bool ignore_checks) {

pos_control->set_max_speed_xy_cm(500);

return true;

}

void ModeCustom::run() {

Vector3f desired_vel(100, 0, 0); // 100 cm/s forward in x

pos_control->input_vel_accel_xy(desired_vel.xy(), Vector2f{0, 0});

pos_control->update_xy_controller();

}

One ArduPilot differentiator: first-class Lua scripting. You can upload a Lua script to the board mid-flight and reshape the mission logic, which is why most agricultural spray solutions (e.g., FlySmart) live on ArduPilot + Lua.

In 2026 the hot ArduPilot areas are (a) underwater ROVs ArduSub — Blue Robotics' BlueROV2 is the de facto kit, (b) precision agriculture with Lua-defined spray patterns, and (c) photogrammetry with the Survey module that auto-generates photo grids.

4. DJI SDK — Mobile / Onboard / Payload, and the US Ban Controversy

DJI still owns more than 70 percent of the global consumer drone market in 2026, and exposes three SDK lines to developers.

- DJI Mobile SDK (MSDK) v5 — Android/iOS camera control, missions, live streaming

- DJI Onboard SDK (OSDK) v4 — runs on the payload bay of the industrial Matrice 300/350 RTK as a ROS/C++ SDK

- DJI Payload SDK (PSDK) v3 — build your own payload (cameras, hyperspectral, LiDAR) and attach it to a DJI airframe

// DJI Mobile SDK v5 (Android) — take a single photo

ActionManager.getInstance().startShootPhoto(

PhotoMode.SINGLE

) { error ->

if (error == null) Log.i("DJI", "Photo taken")

else Log.e("DJI", "Failed: ${'$'}{error.description()}")

}

The hard part is policy. The US NDAA 2025, signed in December 2024, effectively banned DJI components from federal, state, and local procurement starting in 2025. Through early 2026, the conversation has expanded to certain commercial use cases. The ITC's import-ban path is still on hold, but combined with the FAA's mandatory Remote ID and CFR Part 89 rollout, any business with heavy DJI dependence is aggressively diversifying.

Korea, Japan, and the EU did not follow the DJI ban directly, but for any government-funded project the concern is "if we use DJI we cannot enter the US market," which has accelerated self-development or substitution toward Skydio / ACSL.

5. Skydio Autonomy — The US Autonomous Standard

Skydio was founded by MIT alumni in 2014 and, since 2024, has exited consumer drones to focus on enterprise, public sector, and defense. The core is the Skydio Autonomy Engine: six 4K navigation cameras plus an NVIDIA Jetson TX2 (now Orin NX) running real-time SLAM and obstacle avoidance.

The Skydio X10 (2023~) and the defense X10D are programmable through Skydio Cloud + the Skydio Autonomy SDK. For example, automated transmission-tower inspection is roughly an RTSP stream plus a REST workflow:

Skydio Cloud REST — trigger an autonomous inspection mission (pseudo)

token = os.environ['SKYDIO_API_TOKEN']

mission = {

"asset_id": "tower-1234",

"scan_template": "transmission-tower",

"vehicle_id": "X10-SN-9988",

}

r = requests.post(

"https://api.skydio.com/api/v0/missions",

headers={"Authorization": f"Bearer {token}"},

json=mission,

)

print(r.status_code, r.json())

Skydio's strength is being US NDAA-compatible (Blue UAS-listed), with the highest adoption rate inside the US military and law enforcement. The weakness is the price (X10 full kits start around $15,000) and the closed model — you cannot drop into firmware.

6. QGroundControl — The Cross-Platform GCS Standard

QGroundControl (QGC) is a Qt-based ground control station that supports both PX4 and ArduPilot and runs identically on Windows, macOS, Linux, Android, and iOS. With v5.0 in 2026, full MAVLink v2 support, FlySky/Frsky/CRSF RC configuration, real-time video (RTSP/RTP/UDP H.264/H.265), and mission planning (waypoint, survey, corridor) are all standard.

QGC is more than a GCS. You can edit every PX4 parameter (over 2000 of them) from the GUI, and the MAVLink Inspector lets you debug packet-by-packet. If you start a local PX4 SITL, QGC auto-detects it and lets you control the virtual aircraft.

PX4 SITL + QGC (Linux)

cd ~/PX4-Autopilot

make px4_sitl gazebo-classic_iris

Launch QGC in another terminal — it connects on udp:14550 automatically

QGC is dual-licensed open source (GPLv3 + Apache 2.0), and many vendor GCS products are forks. Auterion Mission Control started as a QGC fork.

7. Mission Planner — The Windows-First ArduPilot GCS

Mission Planner is the de facto ArduPilot GCS. Michael Oborne started it solo and still leads maintenance. It is a C# WPF application, so officially Windows-only; macOS/Linux users go through Mono.

Its strength is deep affinity with ArduPilot's internals — parameters, logs, missions. It instantly visualizes .bin logs (ArduPilot's native format), and MAVExplorer / Log Browser draw PID tuning plots. The AutoTune wizard runs end-to-end in the GUI.

| Feature | Mission Planner | QGroundControl |

| --- | --- | --- |

| OS | Windows-first | Cross-platform |

| Firmware | ArduPilot-first | Both PX4 and ArduPilot |

| Log analysis | Full .bin | Full .ulog |

| Mission planning | Strong Survey / Auto Grid | Consistent UI |

| Video | Limited | Rich RTSP/RTP |

| License | GPL | GPL/Apache dual |

Rule of thumb: ArduPilot users pick Mission Planner, PX4 users pick QGC, and mixed shops settle on QGC. The exception is precision-agriculture surveying — Mission Planner's Auto Grid is still faster.

8. MAVLink — The Universal Drone Comm Protocol

MAVLink (Micro Air Vehicle Link) is the comm protocol almost every autopilot speaks: PX4, ArduPilot, parts of DJI, and a Skydio bridge. There is v1 (2009) and v2 (2017), and in 2026 v2 is the standard. It runs over UDP, TCP, serial (telemetry radio), Wi-Fi, and LTE alike.

The key idea is that an XML message definition file auto-generates SDKs in Python, C/C++, Rust, Go, Java, and more.

pymavlink — connect to PX4 SITL over UDP and read HEARTBEAT

from pymavlink import mavutil

m = mavutil.mavlink_connection('udpin:0.0.0.0:14550')

m.wait_heartbeat()

print(f"System {m.target_system}, component {m.target_component}")

while True:

msg = m.recv_match(blocking=True)

if msg.get_type() == 'ATTITUDE':

print(f"roll={msg.roll:.2f} pitch={msg.pitch:.2f} yaw={msg.yaw:.2f}")

MAVSDK (C++, Python, Swift, Java, Go) adds a high-level API on top, maintained by the Dronecode Foundation. A mission such as "take off, climb 5m, move 10m forward, land" becomes a handful of awaited function calls.

MAVSDK-Python — take off to 5m, then land

from mavsdk import System

async def run():

drone = System()

await drone.connect(system_address="udp://:14540")

await drone.action.arm()

await drone.action.takeoff()

await asyncio.sleep(5)

await drone.action.land()

asyncio.run(run())

In 2026 the uXRCE-DDS bridge between MAVLink and ROS 2 DDS has been standardized, so a ROS 2 node can subscribe to PX4-internal uORB topics directly.

9. Pixhawk — Open Hardware Standard (FMU)

Pixhawk is not a single board name but an open hardware spec called the Pixhawk Standards. There are generations FMUv2 (2014) through FMUv6X (2024), and FMUv7 was formalized in early 2026. Anyone can manufacture a Pixhawk board if they meet the spec.

A typical board pairs an STM32H7 (or STM32H7B0) main MCU with triple-redundant IMUs (BMI088, ICM-42688-P, ISM330DHCX), a magnetometer, a barometer, a microSD slot, and USB-C. From FMUv6X onward, dual IMU boards and a separate IO co-processor are part of the spec.

Pixhawk FMUv6X core spec (2024 standard)

- MCU: STM32H753 (480 MHz, Cortex-M7)

- IMU: BMI088 + ICM-42688-P + ICM-20649 (triple)

- Magnetometer: BMM150 / IIS2MDC

- Barometer: BMP388 (dual)

- Interfaces: 8x PWM, 2x CAN, 5x UART, USB-C, microSD

- Power: 5V dual redundant, 3.3V backup

The point of the Pixhawk standard is that firmware (PX4/ArduPilot) is decoupled from boards (vendors). The same PX4 firmware runs identically on Holybro Pixhawk 6X, CUAV X7+, mRobotics, and Auterion Skynode. So a developer is guaranteed firmware portability across boards.

A 2026 trend worth watching is the Pixhawk standard combined with ARK Electronics' ARKV6X — a modular board where you slot in IO co-processors and a Jetson Orin NX so the whole PX4 + ROS 2 + AI vision stack runs on one assembly.

10. Holybro / ModalAI VOXL / NVIDIA Jetson — The Hardware Big Three

When you actually order parts for an open autopilot drone, three vendors dominate.

- Holybro — the default Pixhawk-standard board vendor. The lineup (Pixhawk 6X, 6C Mini, 5X, Durandal) is wide and well-priced (Pixhawk 6X is roughly $250). Direct shipping to Korea and Japan works.

- ModalAI VOXL — successor to Qualcomm's Snapdragon Flight. Compute boards include VOXL 2 (Snapdragon QRB5165 + Hexagon DSP), VOXL 2 Mini, and Sentinel. They support both PX4 and their own Voxl-SDK autonomy stack, and they are US-made.

- NVIDIA Jetson — compute boards. The lineup runs from Jetson Orin Nano (40 TOPS) through Orin NX (100 TOPS) to AGX Orin (275 TOPS). For drones the Orin NX is the workhorse.

Sample 2026 DIY autonomous drone BOM

- FMU: Holybro Pixhawk 6X ($250)

- Compute: NVIDIA Jetson Orin NX 16GB ($699) + Auvidea X220 carrier ($400)

- Camera: Intel RealSense D435i ($349) or Luxonis OAK-D Pro ($499)

- LiDAR: Livox Mid-360 ($999) or Ouster OS0 ($3500)

- Motors/ESC: T-Motor F90 ($35 x4)

- Frame: Tarot 650 ($90)

- Battery: Tattu 6S 6000mAh ($120)

Total: roughly $3500-5500 (depending on choices)

ModalAI is one of the PX4 maintainer companies, so the PX4 integration is the smoothest available. If you want to build something Skydio-like in-house, the standard combo is VOXL 2 + PX4 + voxl-mpa-tools.

11. Microsoft AirSim Sunset (2023) → NVIDIA Isaac Drone Sim

From 2017 to 2023, Microsoft AirSim was the de facto drone/autonomous-driving simulator. Built on Unreal Engine 4 with HIL/SIL bridges for PX4 and ArduPilot, it ran almost like a ROS endpoint via REST and Python. In January 2023, Microsoft announced the end of active AirSim development; the successor is Microsoft Project AirSim, a closed cloud SaaS. The open-source community has kept the lineage alive through Cosys-Lab's Cosys-AirSim fork.

Between 2024 and 2026, the vacancy has been filled by NVIDIA Isaac Drone Sim (the drone-specialized module inside Isaac Sim), built on Omniverse + USD + PhysX 5. RTX ray tracing pushes camera simulation close to photo-realistic.

Isaac Sim Python — spawn a quadcopter into a USD stage

from omni.isaac.kit import SimulationApp

sim_app = SimulationApp({"headless": False})

from omni.isaac.core import World

from omni.isaac.wheeled_robots.robots import WheeledRobot

(For drones, use the dedicated Multirotor extension)

world = World(stage_units_in_meters=1.0)

world.scene.add_default_ground_plane()

world.reset()

for _ in range(1000):

world.step(render=True)

sim_app.close()

Isaac Drone Sim wins on three fronts: (a) Isaac Lab gives you an RL-training environment, (b) Omniverse Replicator auto-generates synthetic datasets, and (c) on a single RTX workstation you can run dozens of drones in parallel.

Alternatives exist — Flightmare (ETH Zurich), Aerial Gym (NTNU), the CARLA Drone branch — but for commercial pipelines, Isaac Sim is becoming the standard fast.

12. Gazebo Sim + ROS 2 + PX4 Integration

Gazebo is the ROS world's original simulator. Gazebo Classic was declared EOL in 2022, and its successor (formerly Ignition) is now just called Gazebo. As of May 2026 the active LTS is Gazebo Harmonic (September 2024 through September 2028).

The integration with PX4 SITL is one command away.

PX4 + Gazebo Harmonic — Iris quadcopter

cd ~/PX4-Autopilot

make px4_sitl gz_x500

Gazebo Harmonic and the uXRCE-DDS bridge come up automatically

Spin up a ROS 2 node and you can subscribe to PX4 topics like vehicle_attitude and vehicle_local_position natively.

ROS 2 — subscribe to position via the PX4 uXRCE-DDS bridge

from rclpy.node import Node

from px4_msgs.msg import VehicleLocalPosition

class Sub(Node):

def __init__(self):

super().__init__('pos_sub')

self.create_subscription(

VehicleLocalPosition,

'/fmu/out/vehicle_local_position',

lambda m: self.get_logger().info(f"x={m.x:.2f} y={m.y:.2f} z={m.z:.2f}"),

10,

)

rclpy.init()

rclpy.spin(Sub())

Gazebo loses on visual fidelity compared with Isaac Sim, but wins on natural integration with the ROS 2 ecosystem. Academic labs still prefer Gazebo, while industry and AI training are migrating toward Isaac Sim.

13. Auterion / AirData — Commercial PX4 Stack and Analytics

Auterion is a Swiss company started in 2017 by PX4 maintainers. They sell "Auterion Skynode + Auterion Suite," a productized PX4. Skynode is a single module that combines a Pixhawk-standard FMU, an NVIDIA Jetson, a 4G LTE modem, and RTK GPS; Auterion Suite is the mission-management and fleet-management SaaS that runs on top.

Auterion's main customers are US defense (US Army Family of Systems - Long Range Reconnaissance), Quantum Systems (German mapping drones), and Freefly Astro (US media drones). It is effectively the OEM standard for NDAA-compatible PX4-based airframes.

AirData is a separate company and a SaaS specialized in flight log (.ulog / .bin) analysis. Upload logs from an SD card or telemetry, and it auto-reports per-motor load, vibration patterns, GPS glitches, and EKF debug traces. Almost every commercial drone operations team uses AirData.

pyulog — extract a PX4 ulog file to CSV (a poor-person's AirData)

pip install pyulog

ulog2csv log.ulg -o ./csv

Produces vehicle_attitude_0.csv, sensor_combined_0.csv, etc.

Companies operating PX4 directly tend to (a) buy the full Auterion Suite SaaS, or (b) combine pyulog plus an in-house data pipeline plus AirData as backup.

14. Companies — DJI / Skydio / Parrot / Yuneec / Autel / Anduril

| --- | --- | --- | --- |

DJI still holds 70 percent or more of consumer and commercial photography in 2026. The Mavic 4 Pro, Air 3S, and Mini 4 Pro lines are still best-sellers. But in the US market they wear an NDAA-noncompliant label that blocks federal procurement, and in 2026 some states are discussing commercial-use restrictions.

Skydio has dropped consumer drones to focus the X10 / X10D plus Skydio Cloud on US government and infrastructure inspection. Parrot effectively cleaned up its consumer line after ANAFI USA and ANAFI Ai and concentrates on government and defense (FreeFlight + Anafi Ai).

Anduril, founded by Oculus founder Palmer Luckey, makes Lattice OS — an operating system for autonomous unmanned aircraft. It builds and co-builds drones like ALTIUS-700M, Bolt, and Bolt-M, plus Roadrunner, a recoverable autonomous interceptor. It is considered one of the companies that has absorbed Ukraine-war data the deepest.

15. Korea — Forest Service + DJI, Self-Development with KARI, Doosan Mobility Innovation (DMI)

Korea's drone industry has three structural features: (a) DJI dependence in the government and public sector is high; (b) a localization track led by KARI (Korea Aerospace Research Institute) is trying to reduce it; (c) Doosan Mobility Innovation (DMI) carves out its own axis with hydrogen-fuel-cell drones.

- Korea Forest Service — operates over 1,000 drones per year for wildfire monitoring, suppression support, and forest surveys. The DJI Matrice 30T was the de facto standard through 2024; since 2025 the service has piloted NDAA-compatible substitutes (Skydio X10, Auterion-Skynode-based domestic airframes).

- KARI Unmanned Aerial Vehicles Lab — develops mid-altitude UAVs like KUS-FS and MUAV for defense and science. On the civilian-collaboration track it supports flight certification for PX4 + Skynode-based industrial drones.

- Doosan Mobility Innovation (DMI) — builds the DS30 hydrogen-fuel-cell multirotor (two-hour flight). Compared with a LiPo drone's 30-minute flight, this is dominant for long-range missions like transmission-line inspection and maritime rescue.

- Hanwha Systems, LIG Nex1 — defense autonomous UAVs. Loitering-munition prototypes informed by Ukraine-war data were unveiled in 2025.

- Coco-Drone, Nearthlab, UCONSystem, Sumbi — civilian autonomy startups. Nearthlab has gone global with autonomous wind-turbine blade inspection.

- MOTIE + MOLIT K-Drone System — a low-altitude UAS Traffic Management (UTM) pilot. In 2026, BVLOS pilot flights began over parts of certain cities.

The Korean market's distinctive feature is that government-led urban BVLOS infrastructure is being deployed quickly. Bundled with UAM, it draws ministry budgets, so the next three to five years should be rich in startup opportunities.

16. Japan — ACSL / ParaZero / Yamaha Agriculture Helicopter / NTT e-Drone Technology

Japan's drone industry resembles Korea's but is different. Japan has (a) more than 30 years of an independent path in agricultural unmanned helicopters, (b) refuses to follow the DJI ban directly but prefers domestic or NDAA-compatible airframes for government procurement, and (c) sees serious drone investments by telcos (NTT, KDDI, SoftBank).

- ACSL — Japanese domestic drone maker. The SOTEN (2021), ACSL-PF2, and PF3-CN are the de facto government-procurement standard. NDAA-compatible; adopted by the National Police Agency, Ministry of Defense, and MIC.

- ParaZero — Israeli headquarters but a heavy Japan presence. Its SafeAir drone emergency parachute is the de facto standard. Japan moves toward mandatory parachutes for urban flight, so ParaZero adoption is very high.

- Yamaha Motor RMAX / FAZER R G2 — agricultural unmanned helicopters that have been in service since 1987. In 2026, more than 40 percent of Japan's agricultural land is still treated with RMAX/FAZER. Note that this is an unmanned helicopter, not a multicopter.

- NTT e-Drone Technology — an NTT Group subsidiary. Covers everything from human-presence services to agricultural spraying to forest mapping. Lineup includes the PRODRONE PD6B-AW-ARM2 and NTT's own PF3-CN.

- KDDI Smart Drone — a KDDI subsidiary. Long-standing 4G/5G cellular BVLOS demos; from 2025 it integrates LTE/5G standard telemetry modules into ACSL drones.

- Terra Drone — surveying and inspection SaaS. Listed in Tokyo and tightly coupled with ACSL.

The Japanese market's distinctive features are the parallel existence of unmanned helicopters (agriculture) and multicopters (photography, industry), and the fact that telcos own BVLOS infrastructure as a business. Korea and Southeast Asia are likely to follow this pattern.

17. Who Should Build Drones — Hobby / Commercial Filming / Industrial / Military

| Scenario | Recommended stack | Why |

| --- | --- | --- |

| Hobby DIY build | Pixhawk 6X + PX4 + QGC + 3-inch cinewhoop | $400-700, rich parts market |

| Consumer photography | DJI Mavic 4 Pro + DJI Fly | Flies out of the box, top camera |

| Commercial filming (IG/YT) | DJI Mavic 4 Pro or DJI Inspire 3 | DJI is optimal outside the US |

| Agricultural spraying | KR - ArduPilot + spray kit; JP - YAMAHA RMAX/FAZER | Depends on acreage and law |

| Forestry / mapping | Auterion Skynode + Holybro + LiDAR + PX4 | NDAA-ready, BVLOS-prepared |

| Wind-turbine blade inspection | Skydio X10 + Nearthlab AI | Autonomy, safety |

| Urban BVLOS delivery | PX4 + ROS 2 + Isaac Sim training | Standard infra rolling out |

| US government procurement | Skydio / Auterion / ModalAI | NDAA compatibility required |

| Military autonomy | Anduril Lattice / PX4 defense forks / Hanwha-class lines | Security and certification |

| Student research | Gazebo + PX4 SITL in sim, then Holybro boards | Lowest cost path |

Three principles. (1) If your target market is the US government, assume NDAA compatibility from day one. Dozens of companies were wrecked by policy in 2025 alone because they leaned on DJI. (2) For self-development, follow the golden path PX4 + Pixhawk + ROS 2 + Isaac Sim — that is where the most learning material lives. (3) Simulate first, fly later. Each crash costs you about $5,000.

18. References

- PX4 Autopilot — https://px4.io/

- PX4 GitHub — https://github.com/PX4/PX4-Autopilot

- Dronecode Foundation — https://www.dronecode.org/

- ArduPilot — https://ardupilot.org/

- ArduPilot GitHub — https://github.com/ArduPilot/ardupilot

- DJI Developer — https://developer.dji.com/

- DJI Mobile SDK v5 — https://developer.dji.com/mobile-sdk/

- DJI Onboard SDK — https://developer.dji.com/onboard-sdk/

- DJI Payload SDK — https://developer.dji.com/payload-sdk/

- Skydio Developer — https://www.skydio.com/skydio-cloud

- Skydio Autonomy — https://www.skydio.com/autonomy

- QGroundControl — http://qgroundcontrol.com/

- Mission Planner — https://ardupilot.org/planner/

- MAVLink — https://mavlink.io/

- MAVSDK — https://mavsdk.mavlink.io/

- Pixhawk Standards — https://pixhawk.org/

- Holybro — https://holybro.com/

- ModalAI VOXL — https://www.modalai.com/

- ARK Electronics — https://arkelectron.com/

- NVIDIA Jetson — https://developer.nvidia.com/embedded-computing

- Microsoft AirSim (archived) — https://github.com/microsoft/AirSim

- Cosys-AirSim fork — https://github.com/Cosys-Lab/Cosys-AirSim

- NVIDIA Isaac Sim — https://developer.nvidia.com/isaac-sim

- NVIDIA Isaac Lab — https://developer.nvidia.com/isaac/lab

- Gazebo (Ignition) — https://gazebosim.org/

- ROS 2 + PX4 (uXRCE-DDS) — https://docs.px4.io/main/en/ros2/user_guide.html

- Auterion — https://auterion.com/

- AirData — https://airdata.com/

- Parrot Olympe — https://developer.parrot.com/docs/olympe/

- Autel Robotics — https://www.autelrobotics.com/

- Anduril Lattice — https://www.anduril.com/lattice/

- KARI — https://www.kari.re.kr/

- Doosan Mobility Innovation (DMI) — https://www.doosanmobility.com/

- Korea Forest Service drones — https://www.forest.go.kr/

- Nearthlab — https://nearthlab.com/

- ACSL — https://acsl.co.jp/

- ParaZero — https://parazero.com/

- Yamaha agriculture helicopter — https://www.yamahamotorsports.com/agriculture

- NTT e-Drone Technology — https://www.nttedrone.com/

- KDDI Smart Drone — https://www.kddi.com/business/iot/drone/

- Terra Drone — https://www.terra-drone.net/