Language models will run everywhere — in the rover, the drone, the robot arm, the sensor on the wall — until every machine can reason about what it is doing, explain that reasoning in plain language, and negotiate directly with people and with other machines. Before acting, an agent will rehearse: running simulations of itself to test a decision, catch its own mistakes and come back with a better plan and explain that reasoning in plain language. We build that future in the open, at Luleå University of Technology, on the open-source ColonyOS runtime.
The lab studies language models as components of control systems, along four themes.
Quantized models run at the asset, which keeps a system working offline and makes natural language the operator interface. The lab spans handheld far-edge devices through on-premise inference with frontier models. Routine decisions run on small models at the asset; the continuum escalates harder reasoning to larger models on-premise or in the cloud.
We study grounding, guardrails and abstention. The open question is not model capability but when an output can be acted on, and decisions must leave auditable evidence.
Digital twins are used both as a development substrate and as a runtime safeguard. Agent-driven experiments are recorded as replayable ColonyOS DAGs.
Workloads are orchestrated across IoT, far-edge, edge and cloud as one system. We study application-aware scheduling and offloading that improve hardware utilization, resilience and predictable performance, where runtime signals from the application let the infrastructure prioritize what matters.
Agents live in the building — reasoning over its digital twin and collaborating with each other and with its occupants to keep it safe, efficient and understood.
Inspection workflows over ultrasonic measurement data, producing auditable reports.
An onboard agent runs mowing missions and answers questions about the robot's own logs.
Three doctoral directions within the lab.
Architectures in which a language model diagnoses and decides while grounding, guardrails and digital twins determine when it must abstain.
Orchestration across IoT, edge and cloud in which runtime signals from the application drive scheduling, scaling and offloading.
Multi-stakeholder systems that keep operating through disconnection and reconcile to one auditable record of who was authorized to do what.
D7065E — Embedded Intelligence at the Edge (7.5 ECTS, Luleå University of Technology). Course notes, tutorials, labs and a building-simulation server are open: github.com/eislab-cps/D7065E.