Flight compute and competing ecosystems
Companion notebook to ModalAI VOXL 2 Mini. That one asks what a single board can do; this one asks what else an indoor autonomous-capture system could be built on, and what an operator could buy off the shelf instead of building one.
Three questions, kept separate because they have different answers:
- Compute — what hardware runs the autonomy stack: integrated autopilot+computer boards, Jetson plus a separate flight controller, or bare Qualcomm modules.
- Turnkey platforms — finished drones and payloads that already capture indoors, where satellite positioning (GPS) is unavailable. The platform side now has its own broader survey — Drone platforms — including camera drones, open research vehicles, and the Chinese ecosystem; this notebook keeps the compute focus and the autonomy scorecard.
- The autonomy question — which of them actually decides where to fly next, as opposed to executing a path an operator drew.
Core take: the compute choice is less interesting than it looks, and the autonomy gap is real. Nothing on the market combines autonomous exploration with photogrammetry-grade imagery — the exploration players are lidar-first and mining-bred, the imagery players are carried by a human.
Prices are July 2026 and moving fast (see the volatility caveat at the end).
Why look past the VOXL 2 Mini at all
Not on merit — notebook 07 concludes the opposite — but because of three practical constraints:
- Lead time. The product page states “Expected to ship within 60 business days” for the Mini, 30 for the full VOXL 2. Roughly three months.
- Userland age. Ubuntu 18.04 / kernel 4.19, ROS 2 (Robot Operating System) Foxy, PX4 1.14. Modern photogrammetry and SLAM (simultaneous localization and mapping) tooling — COLMAP, Open3D, current PyTorch/ONNX — fights that, and neural-processing-unit (NPU) work goes through Qualcomm’s closed SNPE / QNN (Snapdragon Neural Processing Engine, Qualcomm Neural Network) toolchain.
- Licence. Most software development kit (SDK) repos are BSD 3-clause plus a clause restricting use to ModalAI hardware, so nothing lifts onto another board.
Against that, what it uniquely gives is an 11 g complete system with a flight controller, visual-inertial odometry (VIO), mapping and avoidance already integrated.
Integrated compute + autopilot boards
The category VOXL invented and still owns: one board that is both the companion computer and the flight controller (FC).
| Product | Compute | FC | Weight | Price | Buy 1-2? |
|---|---|---|---|---|---|
| VOXL 2 | QRB5165, 15 TOPS | PX4 on DSP | 16 g | $1,299.99 | yes, ~30 days |
| VOXL 2 Mini | QRB5165, 15 TOPS | PX4 on DSP | 11 g | $1,249.99 | yes, ~60 days |
| ARK Pi6X Flow | Pi CM4/CM5, no NPU | ARKV6X | 41 g | $915-1,035 | yes |
| Airvolute DroneCore Jerboa | Orin NX | STM32H7 FMU | 123 g w/ module | quote only | sold out |
| Airvolute DroneCore 2 | Orin NX 8/16 GB | Cube or v6X | 206 g | EUR 2,980 | yes (EU) |
| Sky-Drones AIRLink | RK3399-class, no NPU | yes | 89 g | not published | quote only |
| Auterion Skynode S | quad A53, 2.3 TOPS | FMUv6X | 38 g | not published | no |
Auterion is effectively unavailable outside an OEM relationship. No public prices, no store, Skynode X mission-computer specs unpublished, Skynode N exists only as release notes with no product page, and Skynode Enterprise/OEM is explicitly discontinued. Every path is a sales conversation aimed at OEM volume. AuterionOS is a closed licensed distro on top of PX4/MAVLink. Ruled out for research work on access, not on merit.
Airvolute’s Jerboa is the closest thing to VOXL-class SWaP (size, weight and power) in the Jetson world — 43 g bare, 123 g with an Orin NX 16 GB, integrated STM32H7 flight controller, ~8 W. But it is sold out and quote-only, and a shop page lists 221 g shipping weight against the 123 g spec — unresolved.
Note the weight comparison trap: ModalAI’s 11 g and 16 g are complete systems. Carrier-board weights from ARK, ConnectTech and Auvidea exclude the Jetson module (~28 g) and its heatsink and fan (~52 g). The thermal solution alone outweighs the module.
Terms: TOPS = tera-operations per second (integer inference throughput); DSP = digital signal processor; FMU = flight management unit, the autopilot board proper — microcontroller plus inertial measurement unit (IMU), barometer and magnetometer, as distinct from the carrier board it plugs into; OEM = original equipment manufacturer, i.e. sales only to companies building vehicles in volume.
The split stack: Pixhawk FC + separate computer
The boring option, and probably the right one for bench research.
The Pixhawk world splits the autopilot in two: the FMU carries the microcontroller and the inertial and barometric sensors, and a carrier board provides connectors, power and motor outputs. The Pixhawk Autopilot Bus (PAB) is the standardized connector between them, so the two can be sourced independently. “FMUv6X” and “FMUv6C” name the generation — which microcontroller and which sensor set.
| Board | Standard | Weight | Ethernet | Price | Origin / lead |
|---|---|---|---|---|---|
| Holybro Pixhawk 6X | FMUv6X / PAB | 31.3 g | yes | $166.99 | China |
| Holybro Pixhawk 6C mini | FMUv6C | ~42 g | no | from $130.99 | China |
| ARK ARKV6X | FMUv6X / PAB | 5 g | via carrier | $400.00 | USA |
| Cube Orange+ | CubePilot, not PAB | - | carrier-dep. | $277 | - |
| NXP MR-VMU-RT1176 | FMUv6X-RT | - | yes | $392.70 | 13-week lead |
| mRo Pixracer Pro | FMUv5-class | 8.92 g | no | $349.90 | 3-5 wk lead |
What the Pixhawk standard buys, by design-standard (DS) document: DS-009 fixes JST-GH connector pinouts (a 1.25 mm-pitch locking connector family) so harnesses are cross-vendor; DS-010 is the PAB itself, which makes any compliant FMU drop into any compliant carrier; DS-012 prescribes the v6X sensor layout so PX4 and ArduPilot targets come free. Net effect: swap a $167 Holybro FMU for a $400 US-made ARKV6X without rewiring. That 2.4x price gap is assembly location and defense-procurement compliance, not capability — the same pattern that puts the VOXL at 2.5-3.5x its bill of materials. Cube does not give you this — CubePilot uses its own connector. Ethernet exists only on v6X-class boards, not 6C.
On a short horizon this is the fastest path: a Pixhawk-class FC plus any Linux box gives Ubuntu 24.04, ROS 2 Jazzy, uXRCE-DDS over Ethernet (the current PX4 to ROS 2 transport — micro XRCE-DDS, a thin client for the data distribution service middleware ROS 2 runs on), and an FC that can be replaced within the week. The cost is ~150 g and the integrated VIO/mapping services, neither of which matters on a bench or a slightly larger airframe.
Dead ends worth knowing: Emlid Navio (Emlid now sells only real-time kinematic (RTK) satellite positioning), Aerotenna (defunct, hosting suspended), Aerotenna OcPoC (dropped from PX4), Embention Veronte (EUR 23.5k+, wrong league).
NVIDIA Jetson
The obvious alternative, and it got noticeably weaker in mid-2026.
Modules
Manufacturer’s suggested retail prices at 1,000-unit (1KU) quantity, after the 22 July 2026 increase; specs from NVIDIA. The Orin family is supported to January 2032, none end-of-life.
| Module | TOPS | RAM / bandwidth | Power | Price (old -> new) |
|---|---|---|---|---|
| Orin Nano 4GB | 34 | 4 GB / 51 GB/s | 7-25 W | $199 -> $349 |
| Orin Nano 8GB | 67 | 8 GB / 102 GB/s | 7-25 W | not in the table; post-hike price unverified |
| Orin NX 8GB | 117 | 8 GB / 102.4 GB/s | 10-40 W | $399 -> $649 |
| Orin NX 16GB | 157 | 16 GB / 102.4 GB/s | 10-40 W | $599 -> $999 |
| AGX Orin 32GB | 241 | 32 GB / 204.8 GB/s | 15-60 W | $899 -> $1,799 |
| Thor T5000 | 2070 FP4 TFLOPS | 128 GB | 40-130 W | ~$2,999, pre-increase |
| Thor T2000 / T3000 | 400 / 865 FP4 TFLOPS | 16 / 32 GB | TBA | announced only, Q1 2027 |
FP4 TFLOPS above is tera floating-point operations per second at 4-bit precision — Thor’s headline number, not comparable to Orin’s integer TOPS.
Thor T2000/T3000 are the first genuinely drone-relevant Thor parts — and they ship Q1 2027, with ConnectTech carriers to match. Irrelevant to this project. The shipping Thors draw 40-130 W, which rules them out for a small indoor airframe. Channel pricing is inconsistent this week: Arrow and NVIDIA’s own buy page still show pre-increase numbers.
The software story, which is the real news
Isaac ROS has moved off Orin. Isaac ROS 4.5.0 (July 2026) targets ROS 2 Jazzy on JetPack 7.1, and its supported-platform table lists exactly three things: Jetson Thor (T4000/T5000), x86_64, and DGX Spark. Orin appears nowhere. JetPack 7.2 brings Ubuntu 24.04 to Orin, but NVIDIA staff said in June 2026 that Isaac ROS support is “coming, in the following months”.
So the actual Orin stack today is Isaac ROS 3.2 (Dec 2024) / JetPack 6.2 / Ubuntu 22.04 / ROS 2 Humble — 19 months old. The “modern userland” advantage over VOXL’s Ubuntu 18.04 shrinks a lot.
| Component | Honest grade |
|---|---|
| cuVSLAM (CUDA visual SLAM) | Production-usable. ~116 fps @720p on Orin Nano 8GB, 0.94% translation error on the KITTI driving benchmark. Wrapper Apache-2.0, core is a binary blob. Needs 30 Hz at +/-2 ms jitter — hardware camera/IMU sync is on you, and no ARK/ConnectTech/Auvidea carrier provides it |
| nvblox | Demo-grade for drones. Real TSDF / ESDF / mesh output (truncated and Euclidean signed distance fields — the volumetric map and the collision cost field derived from it), but documented for ground autonomous mobile robots (AMR) with Nav2; the costmap is anchored at the robot pose, which breaks for a vehicle at altitude. No aerial tutorial or benchmark |
| Isaac ROS perception | Genuinely production TensorRT wrappers (YOLOv8, SegFormer, SAM2, FoundationStereo, AprilTag). Nothing aerial-specific |
| PX4 / ArduPilot integration | Pure DIY. PX4’s VIO doc still cites the discontinued RealSense T265, PX4-Avoidance is archived, and ARK’s own VIO->PX4 demo has not been pushed since Feb 2024 |
| ARK-OS | Active, MIT-licensed, but plumbing only: MAVLink routing, RTSP (real-time streaming protocol) video, log upload, DDS bridge. No VIO, no mapping, no avoidance |
There is no NVIDIA drone reference architecture. Both headline reasons to pick Jetson — cheap compute and the Isaac ecosystem — are weaker than they were a year ago.
Jetson carriers
The market has consolidated around the PAB form factor. Weights below exclude the module (~28 g) and heatsink+fan (~52 g). The CSI column counts camera serial interface ports — the direct MIPI camera inputs, as opposed to USB cameras.
| Board | Price | Weight | CSI | Origin | Status |
|---|---|---|---|---|---|
| ARK Just A Jetson | $680 | 75 g | 2x | USA | in stock |
| ARK Jetson PAB V3 | $800 | 66 g | 2x | USA | in stock, but FC breakouts $79.99 not orderable until 1 Sep 2026 |
| ConnectTech Super Hadron-DM | $410 | 45 g | 2x 4-lane | Canada | in stock |
| Auvidea JNX110 | EUR 249-499 | 65 g | 2x | Germany | in stock |
| Airvolute Jerboa | quote | 43 g board | 2x | Slovakia | sold out |
| Holybro Jetson Baseboard | $395.99 | ~190-203 g | 2x 4-lane | China | baseboard only, kits sold out |
| WeAct N006 | $110.50 | 57.8 g | 2x 4-lane | China | AliExpress only |
| Seeed A603 | - | 50 g | 1x | China | 1 camera rules out stereo VIO |
Picks by constraint: CUDA with the least integration work → ARK Just A Jetson (built-in ICM-42688P IMU, 3x UART — universal asynchronous receiver/transmitter, the serial link autopilots use for everything — XT60 battery connector direct to 75 V, no custom harness, ~155 g all-in). European supply with a PAB slot → Auvidea JNX110, which has an integrated STM32F103 running standard PX4 I/O with 8 pulse-width-modulation (PWM) motor outputs. Weight-constrained research → ConnectTech Super Hadron-DM.
Realistic Jetson brain: bill of materials and mass
| Item | $ | g |
|---|---|---|
| ARK Just A Jetson carrier | 680 | 75 |
| Orin Nano 8GB module | ~350-500 | 28 |
| Heatsink + fan | ~40 | 52 |
| ARKV6X FMU | 400 | 5 |
| NVMe solid-state disk, 240 GB | ~60 | 10 |
| RealSense D435i / D455 (both out of stock) | 334-419 | 72-103 |
| CSI stills camera + flat flexible cable | ~50-200 | 15-35 |
| Harness, mounts | ~50 | 20-30 |
| Total | ~$1,960-2,350 | ~280-340 g |
Against a VOXL 2 Mini at $1,249.99 and 11 g for a complete system with a flight controller, VIO, stereo depth and mapping already running. The Jetson brain is roughly 25x the mass and 1.6-1.9x the price, and it hands you no autonomy stack.
Bare Qualcomm modules: a dead end
Buying the same QRB5165 from someone else does not get you what VOXL is.
| Product | Price | Status |
|---|---|---|
| Thundercomm RB5 Core Kit | $795.15 | 2020 platform, aging |
| Lantronix Open-Q 5165RB system on module (SOM) | not published | Last Time Buy — discontinued |
| Qualcomm Dragonwing RB3 Gen 2 (QCS6490, 12 TOPS) | not published | current |
| ADLINK / Excelpoint QRB5165 | - | no current product found |
None of these include a flight controller or an IMU. VOXL’s actual differentiator is PX4 running on the Hexagon DSP with onboard ICM-42688 IMUs; buying a bare SOM means porting that yourself, which is a project in itself. QRB5165 is a 2020 part already in end-of-life motion at SOM level. Do not start new work on it outside VOXL.
Turnkey indoor platforms — the off-the-shelf alternative
| Product | Sensing | Autonomy, honestly | Price |
|---|---|---|---|
| Skydio X10 | vision nav | 3D Scan plans its own path inside an operator-defined volume | ~$16-25k, unverified |
| Skydio R10 | vision, zero-light NightSense | operator-flown, avoidance + backtrack. No mapping product | not public |
| Flyability Elios 3 | Ouster OS0-32 + FlyAware SLAM, caged | vendor says “Not yet” autonomous | ~$25k |
| Emesent Hovermap ST-X | lidar 300 m, Wildcat SLAM | genuine autonomous exploration since Cortex 4.0 | quote |
| Exyn Nexys | modular lidar + fisheye | “Level 4B” — their own scale, not a standard | quote |
| Leica BLK2FLY | lidar nav + scan | autonomous inside a drawn boundary; indoor mode retrofitted 2023 | $44,200 |
| Cleo Dronut X1 | 4K + lidar, ducted | assisted flight | $9,800, dated |
Vendor health matters here. Elios 3 launched in 2022 with no successor announced — a four-year-old platform. Exyn IPO’d in May 2026 raising only $19.4M, against a $75.9M accumulated deficit and $1.19M quarterly revenue (-2.3% year on year); treat as a real vendor risk.
Skydio’s autonomy is not available to third parties. Extend and the Control & Telemetry ICD (interface control document, their published integration spec) expose mission initiation, streaming, telemetry and payload attachment — not the flight-behaviour stack. There is no autonomy SDK.
The competitor that is not a drone
| System | Output | Price |
|---|---|---|
| NavVis VLX 3 | 2x 32-layer lidar + 4x 20 MP cameras, +/-5 mm | quote only |
| Matterport Pro3 | lidar + 134 MP panoramas — the reference navigable tour | $5,995 |
| Leica BLK2GO | handheld SLAM cloud | $60,540 list, $37,550 promo |
| FARO Orbis | SLAM cloud, weak imagery | $19,900 |
| XGRIDS L2 Pro / PortalCam | native 3D Gaussian splatting | EUR 6,295 / 3,799 |
A human with a Pro3 or an XGRIDS handheld is the honest competitor for an occupied building: faster, cheaper and lower-risk than any drone. Any case for an autonomous aerial system has to be made against that baseline, not against other drones.
The autonomy scorecard
The distinction that matters, and that marketing blurs:
- Autonomous exploration — the system decides where to go next. Emesent Hovermap (bounded-volume), Verity and Corvus One (warehouse inventory, lights-out, no GPS, no beacons), Exyn (self-declared).
- Autonomous coverage — path planning inside an operator-drawn volume. Skydio 3D Scan, Leica BLK2FLY. This is what most “autonomous” marketing means.
- Assisted teleoperation — Flyability Elios 3 (vendor admits it), Cleo, Skydio R10.
Verity and Corvus are the proof that indoor autonomy is tractable at scale — neither produces a 3D deliverable. That is the shape of the gap.
Where the gap is
- Nothing combines autonomous exploration with photogrammetry-grade imagery. Exploration players are lidar-first and mining-bred; imagery players are human-carried.
- Room-scale, cluttered, furnished interiors are unserved. Everything autonomous is tuned for large open volumes — mines, warehouses, plants. Doorways, corridors, furniture, glass and mirrors break both SLAM and coverage planning.
- Lighting is the unsolved photogrammetry problem indoors. No vendor claims controlled illumination for image quality; onboard lights are for pilot visibility and produce harsh view-dependent shading that ruins texture consistency.
- No autonomous route to the navigable deliverable. Zillow, CoStar and DJI all shipped Gaussian-splat support in 2025, and XGRIDS sells splat-native handhelds for EUR 4-6k — but capture is still a human walking with a device.
Items 1 and 2 are precisely where the GLEAM / Next-Best-Path line of work sits.
Implications
- On a short horizon, hardware choice is not the bottleneck — and the VOXL ships the VIO and mapping that would otherwise have to be rebuilt. Its 60-business-day lead time makes it an order-early item regardless.
- A split stack is faster to iterate on than any integrated board: a Pixhawk-class FC plus any Linux box gives Ubuntu 24.04, ROS 2 Jazzy, uXRCE-DDS over Ethernet, and an FC replaceable within the week.
- Jetson buys CUDA and an open substrate, not autonomy. With Isaac ROS off Orin and no drone reference architecture, what you get is compute and a build-it-yourself stack.
- The strongest argument against VOXL is roadmap, not capability — see the vendor-direction caveats in notebook 07.
- The differentiating component is the exploration policy, which is hardware-agnostic and portable across every platform here through MAVLink and ROS 2.
Caveats on the figures above
Origin claims are vendor claims. Country of manufacture is taken from vendor pages and is unaudited; the US defense-procurement whitelists that would confirm it are no longer public. It is carried here only because it explains the price spread between otherwise equivalent boards.
Prices are volatile right now. NVIDIA raised Jetson prices 33-101% around 22 July 2026 on memory cost pressure, and distributor channels still show pre-increase figures. Memory pricing also drove LPDDR5X contract prices up ~90% quarter on quarter in Q1 2026. Anything here older than a few weeks should be re-checked.
Specifically unverified, carried forward honestly rather than dropped: Skydio X10/R10 pricing, NavVis pricing, Dronut current pricing, Exyn’s “Level 4B”, Orin Nano 8GB post-increase price, Seeed boxed weights, Auvidea JNX40S/41S/JNX42 dimensions, Airvolute Jerboa price, and whether Auterion Skynode S is Jetson-based. Two open contradictions left visible rather than resolved by preference: Holybro 203.2 g vs PX4 docs 190 g for the same build, and Airvolute Jerboa 123 g spec vs 221 g shop shipping weight.
Figures date from 2026-07-31.