Drone AI Capability Framework - Content Series
Drone AI Capability Framework
Content Series
Draft for Review // April 2026
Series Overview
A multi-post LinkedIn content series that defines AI capability tiers in the drone industry. The series establishes a clear, original framework, then builds on it with editorial takes and deeper analysis.
Positioning goal
Establish Pete as the person who defined AI capability tiers for the drone industry. The framework should be clear and useful enough that others start borrowing it in their own pitch decks and strategy discussions.
Series approach
Post 1 earns credibility by being accurate and useful. Subsequent posts spend that credibility by making claims, asking hard questions, and taking positions.
Planned posts
| # | Format | Working Title | Status |
|---|---|---|---|
| 1 | Carousel (PDF, 8-10 slides) | "Every drone company says they have AI. Here's what that actually means." | In development |
| 2 | Article / long-form | "Why Level 3 drone AI is almost exclusively defense" | Concept |
| 3 | Carousel or article | "The system-of-systems question: does platform autonomy even matter?" | Concept |
| 4 | Article | "What open-source frameworks like dimos mean for NDAA-compliant manufacturers" | Concept |
| 5 | Article or visual | "The AI fluency gap: what the drone industry tells us about AI adoption everywhere" | Concept |
Post 2 takes the editorial position the carousel earns the right to make. Why is Level 3 almost exclusively defense? Is it founder DNA (Silicon Valley tech backgrounds at Shield AI and Anduril), urgency (Ukraine), funding ($20B Anduril contract, $5.3B Shield AI valuation), or something structural about commercial drone markets?
Post 3 addresses the system-of-systems argument (from the Prateek Gupta conversation, April 3 2026): the real "advanced state" isn't removing the human from a single drone but full system integration where a software layer (Lattice, Gotham) connects all assets into a C4ISR system. This is arguably a different axis from platform capability and worth its own post.
Post 4 ties to the dimos/NDAA content idea captured earlier. The full-stack argument for why building on an AI-native foundation beats bolting intelligence onto legacy flight stacks.
Post 5 bridges the drone framework to the broader AI fluency conversation (connects to the Zapier AI Fluency Rubric and NFX "Screenless Startup" ideas captured in #content-ideas). Same underlying pattern: don't bolt AI onto legacy systems, rebuild on an AI-native stack.
Why This Matters
Every drone company says they have AI. But the term means wildly different things depending on who is talking. A company running YOLO-based object detection on an FPV feed calls it AI. A company whose drone autonomously flies simulated F-16 dogfights in GPS-denied environments also calls it AI. For C-level decision-makers evaluating drone capabilities, this ambiguity is a real problem.
No widely adopted framework exists that maps AI capability levels specifically to the drone industry. The automotive world has SAE J3016 (Levels 0-5 for self-driving). The defense world uses "human-in-the-loop / on-the-loop / out-of-the-loop" language. But nobody has combined these concepts into a drone-specific taxonomy with real company examples that a buyer or investor can use.
Background and Research
Origin of the idea
The concept started with Jensen Huang's discussion of perception AI on the Lex Fridman podcast. The observation was that many drone companies are presenting "AI-powered" solutions that are really just YOLO-based object detection layered onto existing hardware. Meanwhile, defense companies like Shield AI and Anduril are building systems that make genuine autonomous decisions. The gap between these two is enormous, but the industry uses the same word ("AI") to describe both.
Existing frameworks considered
- SAE J3016 (automotive): Levels 0-5 for driving automation. Well-known but specific to vehicles. Zipline borrows "Level 4" to describe their drones, but it is not an industry-wide standard for UAS.
- Human-in/on/out-of-the-loop (defense): Originated from a 2012 Human Rights Watch report. Palmer Luckey and Anduril use this language publicly. Anduril describes their systems as "human-on-the-loop." Well-established in defense circles but generic and not drone-specific.
- NVIDIA perception/physical/agentic AI: Jensen Huang's framing for the broader AI/robotics progression. Useful conceptually but not mapped to specific drone capabilities or companies.
Key research finding
The original thesis was that "agentic drones are the sole domain of defense companies." Research largely confirms this with one important nuance: defense companies are ahead on adversarial agentic AI, where drones make decisions while someone is actively trying to defeat them (GPS jamming, comms denial, contested airspace). Commercial companies like Zipline and Skydio are achieving high levels of autonomy, but in cooperative environments. This is a fundamentally different problem. Level 3 capability, as currently defined, is almost entirely Anduril and Shield AI.
Post 1: The Framework Carousel
Goal
Publish a LinkedIn carousel that establishes the framework descriptively. Map real companies to each tier. Position this as a vertical-agnostic landscape map so that readers across defense, commercial, public safety, and industrial verticals can all see how the framework applies to their world.
The Framework
Built around one axis: who is making the decision, and when. This maps cleanly to both the "human loop" spectrum and the AI capability required. The framework is vertical-agnostic: it describes what the AI on the platform can do, regardless of whether the drone is inspecting a power line, delivering a package, or flying a combat mission.
AI Capability Level 1: The drone sees, a human decides
The AI handles perception: object detection, classification, tracking. Every meaningful decision (where to go, what to do, when to act) is made by a human operator. The drone is essentially a smart sensor. Most of the industry lives here today, and many companies marketing "AI-powered" drones are describing Level 1 capability.
AI Capability Level 2: The drone sees and acts within boundaries a human set
The AI handles navigation, obstacle avoidance, mission execution, and real-time adaptation (rerouting around weather, adjusting scan patterns). The drone operates within a mission plan or set of rules that a human defined beforehand. It does not decide *what* to do. It decides *how* to do what it was told.
The jump from Level 1 to Level 2 is about navigation and execution autonomy.
AI Capability Level 3: The drone sees, reasons, and decides what to do next
The AI makes decisions about goals, not just execution. It can re-prioritize, retask itself, and respond to novel situations it was not explicitly programmed for. In defense, this means tactical decision-making under uncertainty with no comms link. In commercial, this would mean a drone that discovers an anomaly during inspection and autonomously decides to investigate further, log it, and adjust the rest of its mission.
The jump from Level 2 to Level 3 is about decision-making autonomy, which is a qualitatively different capability.
Mapping to existing frameworks
| Framework | Level 1 | Level 2 | Level 3 |
|---|---|---|---|
| Human loop | Human-in-the-loop | Human-on-the-loop | Human-out-of-the-loop |
| SAE equivalent | L1-L2 | L3-L4 | L4-L5 |
| NVIDIA framing | Perception AI | Physical AI | Agentic AI |
Company Mapping
Companies are placed based on their most capable shipping product or system. No straddling; each company gets one tier.
*Preliminary. Full company research (JAM-17) will validate and expand this list.*
| Level | Companies (cross-vertical) |
|---|---|
| Level 1 | DJI, Autel, Brinc, FLIR/Teledyne, DroneUp, DroneSense, most DFR vendors |
| Level 2 | Skydio, Zipline, Matternet, Wing (Alphabet), Percepto, Flytrex |
| Level 3 | Shield AI (Hivemind), Anduril (Lattice + Fury/Altius) |
Carousel Slide Outline
Format: PDF carousel, 1080x1350px (portrait), 8-10 slides. One idea per slide. Large text, minimal clutter. Clean design inspired by Ben Evans landscape maps.
| Slide | Visual | Text | Takeaway |
|---|---|---|---|
| 1 (Hook) | Bold text on clean background, no logos | "Every drone company says they have AI. Here's what that actually means." | Stop the scroll. Create curiosity gap. |
| 2 (Problem) | Pattern interrupt: different background color. Simple stat or provocative line. | "There are 3 distinct levels of AI capability in drones. Most of the industry is at Level 1." | Establish that "AI" is not one thing. Most companies are at the bottom. |
| 3 (The axis) | Simple horizontal spectrum graphic: Human decides → Human sets boundaries → Drone decides | "The framework: who is making the decision, and when." | Introduce the organizing principle before the tiers. |
| 4 (Level 1) | Level 1 label + one-line definition. Company logos below. | "Level 1: The drone sees, a human decides. AI handles perception (detection, classification, tracking). Every real decision is human. This is where most 'AI-powered' drones live today." | Reader can immediately benchmark companies they know. |
| 5 (Level 2) | Level 2 label + one-line definition. Company logos below. | "Level 2: The drone sees and acts within boundaries. AI handles navigation, obstacle avoidance, mission execution. It decides *how*, not *what*." | The jump is execution autonomy. |
| 6 (Level 3) | Level 3 label + one-line definition. Company logos below. | "Level 3: The drone reasons and decides what to do next. AI makes goal-level decisions. Re-prioritizes, retasks, handles novel situations without a human in the loop." | This is where it gets real. Very few companies are here. |
| 7 (The map) | Full landscape map: all three tiers with company logos placed. Clean grid or spectrum layout. | "The drone AI landscape, mapped." | The shareable slide. People save and screenshot this one. |
| 8 (The gap) | Visual emphasis on the gap between Level 2 and Level 3. Few logos on the right side. | "Notice the gap. Level 3 is almost entirely defense. Why? That's the question that matters." | Tease the follow-up post. Create anticipation. |
| 9 (CTA) | Clean slide with Pete's name/photo and a clear call to action. | "Follow for the next post in this series: why Level 3 drone AI is almost exclusively defense, and what that means for the rest of the industry. 💬 Where would you place your company?" | Drive engagement (comments), follows, and saves. |
Distribution and Tagging Conventions
Company mentions: Tag the company's LinkedIn page for every company on the map. This increases the chance company employees see and reshare. Prioritize tagging companies where Pete has a connection who works there.
Individual mentions: Tag specific people who would find this relevant or who might amplify it:
- Founders/CEOs of mapped companies (especially smaller ones who are more likely to engage)
- Industry analysts and journalists who cover UAS
- People Pete has had recent conversations with about this topic (e.g., Prateek Gupta)
Hashtags: Use 3-5 targeted hashtags. Avoid generic (#AI, #drones). Prefer:
- #UAS
- #DroneIndustry
- #AutonomousSystems
- #DefenseTech
- #DroneAI
Caption: The carousel gets a short text caption (2-3 sentences max) that frames the post and asks a question to drive comments. Example: "Every drone company says they have AI. But there are really 3 distinct capability levels, and most of the industry is at Level 1. I mapped the landscape. Where does your company fit?"
Timing: Post Tuesday-Thursday, 7-9am PT for maximum feed visibility.
Open Questions
- Naming: "AI Capability Level 1/2/3" is functional but not memorable. The final terms should be short, plain, and immediately obvious. Options include coining original terms or adapting existing language. The names should be intuitive enough that a CEO reading a slide gets it without a legend.
- Company coverage: The current list is incomplete. JAM-17 research will expand and validate. Companies like Redcat, Flightwave, Flock Safety, Dedrone, senseFly/AgEagle, AeroVironment, and others need to be researched and categorized.
- Level 2/3 boundary: The framework needs a clear litmus test for when a system crosses from "executing a plan intelligently" to "making its own plan." Proposed test: can the system autonomously change its mission objective (not just its route or method) based on new information, without human approval?
- Design: Need to decide on visual style, color palette, and whether to use actual company logos (permission considerations) or company names in text.
Next Steps
- Conduct exhaustive company research across verticals to build a complete company list (JAM-17).
- Categorize each company against the three-level framework based on most capable product.
- Pressure-test the framework: do the categories hold? Are there companies that break the model?
- Finalize tier names.
- Design carousel slides (clean, simple style inspired by Ben Evans).
- Draft carousel caption and compile tag lists (companies + individuals + hashtags).
- Publish and monitor engagement.