Concept 1: What is Edge AI?
Edge AI means running artificial intelligence locally on a device rather than sending data to a remote cloud server.
Why it matters:
| Cloud AI | Edge AI |
|---|---|
| Requires internet | Works offline |
| Slower (data travels far) | Faster (processes locally) |
| Privacy concerns | Data stays on device |
| Ongoing API costs | No cloud fees at runtime |
Real example from the article:
The Reachy Mini Jetson Assistant runs "everything locally with GPU acceleration — no cloud, no API keys, no internet required at runtime"
Concept 2: What is NVIDIA Jetson?
NVIDIA Jetson is a family of compact computing modules designed specifically to run powerful AI workloads at the edge.
Think of it this way:
Traditional AI Setup:
[Data] → [Internet] → [Server Room] → [Result]
Jetson Setup:
[Data] → [Jetson Module] → [Result]
(fits in a handbag)
Key characteristics:
- Compact — physically small enough to embed in robots
- Powerful — capable of running frontier AI models
- Complete — includes the full AI software stack
- Portable — battery-friendly for real-world deployment
Concept 3: Understanding AI Performance — What is TOPS?
TOPS stands for Trillion Operations Per Second — it measures how fast an AI chip can process calculations.
Simple analogy:
Think of TOPS like a chef's speed in a kitchen. More TOPS = more dishes prepared per second
Jetson lineup by TOPS:
Jetson Orin Nano Super → 67 TOPS (Beginner)
Jetson AGX Orin → 275 TOPS (Intermediate)
Jetson AGX Thor → 2070 TFLOPS (Advanced/Professional)
Rule of thumb: More complex AI tasks require more TOPS
Concept 4: The Jetson Product Tiers — Matching Power to Purpose
The Jetson platform follows a progressive tier system, letting builders start small and scale up.
Tier 1: Jetson Orin Nano Super
Best for: First-time builders, students, hobbyists
- 67 TOPS of AI performance
- Entry-level price point
- Ideal use cases:
- ✅ Learning computer vision
- ✅ Building first AI agents
- ✅ Prototyping edge AI projects
Real projects built with it:
- A toy car that drives itself autonomously (SidewalkPilot)
- A voice + vision assistant running fully on-device
- A first AI robot built from scratch using open-weight models
Tier 2: Jetson AGX Orin
Best for: Advanced makers, researchers, startup prototypes
- 275 TOPS of AI performance
- Handles more complex, simultaneous workloads
- Ideal use cases:
- ✅ Advanced robotics curriculum
- ✅ Autonomous navigation
- ✅ Industrial automation
- ✅ Smart vision systems
Real projects built with it:
- A live browser-based Vision Language Model interface
- An autonomous rescue robot that maps environments in real time (Carnegie Mellon University)
Tier 3: Jetson AGX Thor
Best for: Professional researchers, humanoid robotics, enterprise
- 2,070 FP4 teraflops + 128GB memory
- Server-class compute in a portable form
- Ideal use cases:
- ✅ Humanoid robots
- ✅ Real-time sophisticated reasoning
- ✅ Factory and field deployment
- ✅ Multimodal AI (vision + audio + language combined)
Real projects built with it:
- Two robotic arms that autonomously prepare matcha (hackathon winner)
- A multimodal AI studio combining vision, audio, and language
Concept 5: What is a Vision Language Model (VLM)?
A VLM is an AI model that can see and understand images/video and respond in natural language.
How it works:
[Camera Feed] ──→ [VLM] ──→ [Text/Action Response]
"What do you see?" "I see a person waving"
Why it matters for robotics:
Robots need to perceive their environment and reason about it — VLMs enable exactly this, combining:
- 👁️ Vision — understanding what the camera sees
- 💬 Language — describing, reasoning, and responding
Concept 6: Agentic AI at the Edge
AI Agents are systems that can plan, decide, and act autonomously to complete goals — not just answer questions.
Reactive AI vs. Agentic AI:
| Reactive AI | Agentic AI |
|---|---|
| Responds to input | Pursues goals independently |
| Single task | Multi-step reasoning |
| Passive | Takes real-world actions |
On Jetson specifically:
The article mentions "Jetson Device Skills" and "Jetson BSP Skills" — tools that help developers deploy coding AI agents that can:
- Create solutions
- Optimize performance
- Deploy to real-world edge environments
Concept 7: The Full AI Stack
The article mentions Jetson gives developers "the entire AI stack." Here's what that means:
┌─────────────────────────────┐
│ Applications/Robots │ ← What users interact with
├─────────────────────────────┤
│ AI Models (VLMs, │ ← The intelligence layer
│ LLMs, Vision Models) │
├─────────────────────────────┤
│ AI Frameworks & Tools │ ← Software to run models
│ (NVIDIA Isaac, GR00T) │
├─────────────────────────────┤
│ GPU Hardware │ ← Physical compute power
│ (Jetson Module) │
└─────────────────────────────┘
Having the full stack means developers don't need to piece together separate tools — everything works together out of the box.
Summary: How All Concepts Connect
EDGE AI (run AI locally)
↓
JETSON PLATFORM (the hardware that enables it)
↓
TOPS/TFLOPS (measure how powerful each module is)
↓
THREE TIERS (Nano Super → AGX Orin → AGX Thor)
↓
VLMs + AGENTIC AI (the intelligence running on the hardware)
↓
FULL AI STACK (everything integrated and ready to deploy)
↓
REAL-WORLD ROBOTS & AUTONOMOUS MACHINES
Quick Decision Guide: Which Jetson is Right for You?
| Your Situation | Recommended Module |
|---|---|
| First robotics project, student | Jetson Orin Nano Super |
| Advanced research, startup prototype | Jetson AGX Orin |
| Humanoid robots, enterprise, cutting-edge research | Jetson AGX Thor |
💡 Key takeaway: Start with the tier that matches your current needs — the platform is designed so you can upgrade progressively as your projects grow in complexity.