> Markdown version of [/events/world-congress-2026-europe/sessions/1324-from-perception-to](https://www.wearedevelopers.com/events/world-congress-2026-europe/sessions/1324-from-perception-to). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

# From Perception to Autonomy: Building Agentic Edge AI Robots with ROS 2


- **Date:** Friday, Jul 10, 2026
- **Time:** 13:40–14:10 (30 min)
- **Room:** Stage 5
- **Event:** World Congress 2026 Europe

## Recording

[Watch recording](https://www.wearedevelopers.com/videos/100295-from-perception-to-autonomy-building-agentic-edge-ai-robots-with-ros-2)

## Description

AI in robotics is moving fast — from cloud-dependent pipelines to fully autonomous, on-device systems that perceive, reason, and act without connectivity. But for most developers, the gap between "cool demo" and "working robot" is still wide. This talk bridges that gap.
I'll walk you through the evolution of Physical Edge AI — what it actually means in practice, not just in keynotes — and show you how to build a real autonomous system using ROS 2, Edge Impulse, and Qualcomm edge hardware.
The centrepiece is a live architecture walkthrough of an autonomous inspection robot I built and demoed at Embedded World 2026. It runs two AI models on-device: a fast object detector to decide where to look, and a close-up anomaly scorer to decide is this a problem? — all orchestrated as ROS 2 nodes on a Qualcomm QCS6490 (RubikPi 3), with zero cloud dependency.
From there, I'll zoom out to what's coming next: agentic edge AI — robots that don't just detect and classify, but plan, decide, and adapt using lightweight LLMs/VLMs running locally on edge hardware. I'll cover the practical building blocks available today (TensorRT Edge-LLM, quantised VLMs on Jetson, vision-language-action models), the ROS 2 integration patterns that make this work, and the honest gaps that still need solving.
What you'll take away:

A reusable architecture pattern for multi-model edge AI robotics (detect → inspect → decide)
How to integrate Edge Impulse ML models as native ROS 2 nodes
What "agentic" actually means at the edge — and what's real vs. hype in 2026
Practical guidance on edge hardware selection (Jetson, Qualcomm, NPUs) and when to go hybrid edge-cloud
Where the open-source ecosystem (ROS 2, Edge Impulse, NVIDIA Isaac) is headed for developers who want to build physical AI systems

## Speaker

### [Moe Sani](https://www.wearedevelopers.com/@moe-sani)

Edge AI Solutions Architect at Qualcomm


## Related talks at this congress

- [Edge AI on iOS: Beyond the Cloud, Designing the Next Generation of Intelligent On-Device Apps](https://www.wearedevelopers.com/events/world-congress-2026-europe/sessions/1256-edge-ai-on-ios) — Precious Osaro
- [Ducks, Sensors & Agents: Hands-On Edge AI with Arduino UNO Q](https://www.wearedevelopers.com/events/world-congress-2026-europe/sessions/1145-ducks-sensors-agents) — Jenny Speelman, Dmitry Maslov
- [How I built my own intelligent Robot Arm from Scratch](https://www.wearedevelopers.com/events/world-congress-2026-europe/sessions/1090-how-i-built-my-own) — Iulia Feroli
- [Edge Orchestration for the Physical World: Connecting Cameras, Sensors, and Devices with MQTT](https://www.wearedevelopers.com/events/world-congress-2026-europe/sessions/1084-edge-orchestration) — Irina Terekhova