Overview
This use case demonstrates how RIoT Secure enabled artificial intelligence to run directly on highly constrained embedded hardware in a safety-critical environment, without relying on cloud connectivity or additional infrastructure.
In collaboration with Neuton.ai, a provider of ultra-compact neural networks (now part of Nordic Semiconductor), RIoT Secure supported the development and deployment of an edge AI solution at Stockholm Arlanda Airport on baggage handling vehicles, where it was used by SAS (Scandinavian Airlines) and Swedavia (the Swedish airport operator) to improve safety in airport ground operations by reliably detecting when vehicles entered or exited indoor areas.
The Challenge
Airport ground vehicles are required to automatically reduce speed when entering indoor environments such as baggage handling areas. At Stockholm Arlanda Airport, located at approximately 59 degrees latitude, this requirement proved difficult to enforce using traditional GNSS-based geofencing.
GNSS accuracy degrades significantly at high latitudes due to satellite geometry and limited visibility. In practice, positional errors of 10–20 meters were common, particularly near buildings and indoor transitions. This made conventional GPS geofencing unreliable and, in many cases, unusable. Incident data showed that vehicles could exceed speed limits indoors without detection, contributing to serious safety incidents. Cloud-based solutions were not viable due to connectivity limitations in indoor environments and the need for immediate, deterministic responses. Installing additional infrastructure such as beacons was also undesirable in a busy airport setting.
The challenge, therefore, was not location tracking, but reliable classification of indoor versus outdoor operation, under strict hardware, latency, and operational constraints - in an environment where failure carries a real safety risk.
The Solution
The solution was built around a strict separation of concerns that allowed artificial intelligence to be introduced into an existing, production-grade embedded systems without disrupting established safety, security, or operational functionality. Rather than attempting to improve absolute positioning, the AI approach focused on interpreting raw GNSS signal characteristics to infer environmental context. Specifically, the approach analyzed GNSS GSV (satellites-in-view) data, including satellite visibility, signal-to-noise ratios, and signal geometry, to determine whether a vehicle was operating indoors or outdoors.
Neuton.ai delivered the indoor/outdoor detection capability as a compact, C-based classification library, designed to run efficiently on highly constrained hardware. This made it straightforward to integrate the AI logic directly into the existing application code running on an ATmega2560 (8-bit MCU), without requiring architectural changes, additional dependencies, or modifications to real-time behavior. The microcontroller remained fully dedicated to data processing, inference, and decision-making, including control of a relay that enabled or disabled vehicle speed limitation when entering indoor areas. The AI-based GNSS signal analysis was introduced incrementally, running in parallel with already deployed services and building on RIoT Secure’s existing production deployment with SAS. These services included local safety and security functions, operational data collection, and enforcement of driver badge registration. The new AI functionality operated alongside these components as an additional application capability, rather than as a replacement or refactor of existing logic.
RIoT Secure independently managed secure communication, firmware delivery, and device lifecycle operations through its platform using a separate Arduino MKR GSM 1400 or Ardunio MKR NB 1500, utilizing a secure element, to ensure that lifecycle management concerns remained completely isolated from application execution. This allowed AI specialists to focus solely on model behavior and signal classification, while RIoT Secure ensured that updates, connectivity, and long-term device governance could evolve safely over time. By maintaining clear boundaries between application logic and lifecycle infrastructure, the system was able to incorporate new AI-driven behavior without increasing operational risk - even in a safety-critical airport environment.
Why This Use Case Matters
This deployment demonstrates that meaningful AI at the edge does not depend on powerful hardware, continuous connectivity, or cloud-based intelligence. Instead, it depends on architectures that are designed to operate reliably under real-world constraints.
By running AI-based GNSS signal classification directly on an ATmega2560, the system was able to make safety-critical decisions in real time, even in environments where traditional positioning techniques fail and connectivity cannot be assumed. The solution avoided additional infrastructure, operated independently of the cloud, and delivered deterministic behavior in a safety-sensitive airport setting. Equally important, the AI capability was introduced incrementally into an existing production system. New intelligence was added alongside established safety, security, and logistical services without disrupting operational workflows or increasing lifecycle risk. This demonstrates how edge AI can evolve over time, rather than requiring disruptive redesigns or one-off deployments.
More broadly, this use case highlights a fundamental requirement for deploying intelligent systems at scale:
IoT lifecycle management must be treated as infrastructure - not as an application responsibility.
When communication, security, and updates are isolated from application execution, AI teams gain the freedom to innovate, while operators retain control over long-lived devices in the field. This approach enables organizations to deploy intelligent functionality on constrained, or physically inaccessible devices while continually improving such systems safely over their entire operational lifetime.
The AI application ran on an 