Master's Thesis · DSV · Spring 2026
Semantic Vision-Language reasoning for autonomous lawn mowers.
A modular split-compute architecture that pairs lightweight onboard segmentation with edge-deployed VLM reasoning · turning what a Husqvarna Automower sees into structured natural-language decisions about what to do next.
STAGE 1 · ONBOARD SEGMENTATION
PERSON · 0.92
PERSON · 0.88
Edge VLM · structured outputidle
Final decision0.91
awaiting decision
Route away from the occupied picnic area instead of passing close to the people.
Onboard: Raspberry Pi 4 · segmentation→Edge: VLM node · semantic reasoning
Fig. 01 · Test frame · people-having-a-picnic scenario
01 · Overview→
The research question, the limits of fixed-vocabulary perception, and what semantic reasoning unlocks.
02 · Architecture→
Onboard segmentation on a Raspberry Pi, edge-deployed VLM, and a structured natural-language contract between them.
03 · Experiments→
Six outdoor scenarios on real robot footage, dual-frame inference, and a 18-frame curated benchmark.
Conducted in collaboration with
HusqvarnaRI.SEStockholm UniversitySupervisors · Fehmi Ben Abdesslem · Miriana Passarotto · Xiaodan Shi