A drone interpreting an image cannot always wait for a remote server’s answer. Yet adding onboard processing also brings constraints. Recent research examines this balance for civilian applications.

Three architectures compared

The preprint Cloud, Edge, or Split?, submitted on 21 September 2026, examines fully onboard, remote and split processing. Zoha Azimi, Reza Farahani, Schahram Dustdar and Christian Timmerer use SmolVLM-256M as a representative lightweight vision-language model.

Connectivity is part of the system

The ATHENA laboratory also presented the research in September. The researchers measure processing delay, computing resources, data traffic and energy across different images and network conditions. Their conclusion: none of the three strategies is universally optimal.

A result with a limited scope

This is a profiling study of a lightweight model, rather than flight certification or a demonstration of general autonomy. Another model, aircraft or connection may behave differently. Interpretation quality and processing speed remain separate criteria.

Implications for civilian inspection

Our analysis: roof or infrastructure inspection must also account for connectivity interruptions, image privacy and human supervision. Local processing can limit transmissions; remote processing provides resources but introduces another dependency.

Conclusion: match the architecture to the mission

The issue extends beyond processor power. The architecture must fit the conditions actually encountered. This study reminds us that credible onboard AI is assessed across the whole service, beyond an isolated demonstration on an ideal network.