Current state of drones, commercial + DoD: Conversation with a friend on 2024-12-05

Will relay what I know. I will add a SME (subject matter expert) level. 100% being highest confidence in what I know, 0% being lowest.

I will add more during our conversation and update the website accordingly.

Hardware

Payload

SME level: 80%

Industry moving toward Low-SWaP (size, weight, and power) hardware, specifically for DoD applications.

High-SWaP is good for a "mother ship", with multiple low-SWaP drones it communicates with.

• nvidia Jetson/Orin/AGX/etc.
  • Uses TensorRT to "compile" a computer vision DL models (either detection, classification, tracking, etc.) specified to the hardware of the inferencer, as well as numeric precision of choice.

Sensors / Gimbals

Can briefly discuss these. I don't have exact product links on-hand.

Some gimbals are "smart". Meaning, given a GPS/GNSS antenna connected to it (or, most have one built-in), you can say "look at this Lat/Lon", and given the camera intrinsics + extrinsics, it can figure out where how much to pan/tilt to center that position. Some can have built-in tracking.

EO

SME level: 20%

MSI

SME level: 50%

HSI

SME level: 50%

LIDAR

SME level: 0%

SAR (Synthetic Aperture Radar)

SME level: 60%

Needs a good IMU for good image-formation.

Vendors:

• IMSAR
  • They make their NSP series. Smaller number = smaller radar.
  • Have an internal IMU within the NSP, highly sensitive.

Platform (Group 2 UAV's)

SME level: 25%

Can be fixed-wing or rotorcraft. Many DoD applications use fixed-wing. Can also be hybrid, with VTOL.

• Aerosonde by Textron - good, interesting land/takeoff
• ScanEagle by Boeing - old
• Puma by AeroVironment - proprietary controls
• Stalker by Lockheed Martin - VTOL

Platform (Group 1, 3+ UAV's)

SME level for Group 1: 10% SME level for Group 3+: 0%

Group 1 can include your hobby drones. Group 3+ are large, outside the scope of this conversation.

Comms

General

SME level: 100%

Stick with encrypted radio comms + UDP for christs sake. Nothing with TLS, nothing like HTTP or TCP. I have seen this mistake so many times.

GCS (Ground Control Station) Radio

SME level: 5%

• Lower freq = longer range + smaller data rate
  • This is usually for very long-range UAV's
• Higher freq = shorter range + larger data rate
  • This tends to be for FPV drones

Usually around 1.5-2.5 GHz is sweet spot for group-2 UAV's.

They always encrypt the packets, but always double-check.

Vendors:

• Silvus
  • Tend to operate anywhere from mid (~500) MHz range -> 6GHz.
  • Is usually mesh network.
  • Omni-directional.
• Moog
  • Don't know the exact models, but Moog is a general USG / DoD contractor.
  • The most experience I had were directional antennae, and they were amazing.
• There are more

UAS (Unmanned Aerial System) Radio

SME level: 0%

Literally don't know.

Standards

SME level: 100%

• MISB
  • NGA + National Standards
  • Tooling for MISB (made by me): misb Rust crate
  • This standard is GOOD. It is packet-ized, low bandwidth, and easy to use.
  • To make own format using KLV (key-length-value), see (made by me): tinyklv Rust crate
• STANAG
  • NATO STANdardization AGreement. Has many forms, a couple of which pertain to UAS comms.
  • E.g. STANAG 4676 - tracking
  • E.g. STANAG 4586 - general (see page 16/17 for more STANAG standards)
  • The standard itself is good, however the format is bound by XML. This means, you have to serialize your information into plain-text XML and send it down a data-link. This is atrociously slow.
  • Can still use the standard, but with a different format, e.g.: KLV.
• CoT (Cursor on Target)
  • MITRE standard, useful for detection, classification, and tracking.
  • Also XML bound. Is useful for displaying information.

Algorithms

SME level: 100%

I won't say examples here. There are a LOT. But can discuss them - AAV

Note that these all depend on your modality of data. E.g. SAR can see through moisture and many atmospheric artifacts, but is very rough (visually) in its current state for UAV's (turbulence, de-speckling, filtering, all factor a role). In addition, MSI/HSI can do more refined identification rather than classification (e.g. "it is ship A" rather than "it is a ship").

All are best when done on the edge. Downlink of data and then processing helps with compute, but you are susceptible to latency on decision making, compression artifacts, etc.

• Detectors
• Classifiers
• Detector-classifiers (combined)
• Trackers
  • Accounting geodesic information for 3D, world-space, tracking. Rather than pixel-based
• Path planning (platform)
  • Accounting for no-fly-zones
  • For some sensors like SAR, a specific geometry is required for the platform to maneuver to capture the SAR image.
• Path planning (gimbal)
  • Planning where the gimbal is pointing. E.g. wide-area-search
• Swarming algorithms
• World-state awareness
• Global self-awareness
  • This means awareness of the platforms current extrinsics, velocity, how much fuel, capabilities, etc.
• Reasoning and prioritization
  • Prioritization of tasks to perform
  • Reasoning on how to perform each task
• Natural language command + control

Development

SME level: 90%

• mosh: stands for "mobile shell". If you need to access a computer and make quick changes, usually use SSH. But that requires a persistent connection. With drones + radio, your comms are always in/out. mosh helps by keeping the session persistent, and sending back screen-caps of what you are doing.
• Python, TensorRT, ONNX, etc. - Deep Learning Tools
• Embedded software - Rust (good), C++ (good IFF expert, otherwise unsafe)

General comments

One thing that people often ask about is {insert company name here}. I'm not typing it because I don't want people to find it when searching. But this company has absolutely horrendous hardware, and absolutely horrendous software FOR "mission autonomy". However, they do have good "command + control" software.