A reasoning engine in every machine
Local LLMs on limited hardware, at the asset: offline, sovereign, low-latency. Talk to any device in plain language; the machine reasons over its own sensors, logs, and self-model.
StyrOps CPS Lab
Self-reasoning
Self-reasoning
cyber-physical systems
Machines that carry their own reasoning, improve themselves like scientists, and are born in simulation. A research lab at Luleå University of Technology, built on the open ColonyOS substrate.
Research vision
The LLM becomes a first-class component of cyber-physical system design.
Improvement as empirical engineering
Self-improvement is not blind patching. Agents form hypotheses, run controlled experiments, evaluate versioned oracles, and pass guards — every step an immutable, replayable process chain.
Born and verified in simulation
Agents generate digital twins of the systems they will run, develop and verify against the twin, and keep it in the loop as a guardrail oracle once deployed.
Research challenges
Agentic AI on real assets is not a solved engineering problem. The open questions are not about model size — they are about what it takes for an answer to be acted on.
Benchmarks are not plants
Agents that ace benchmarks still fail on real assets. Live systems bring messy signals, missing context and consequences that no leaderboard measures.
The question is not “can it answer?”
On a safety-critical system the question is whether the answer can be trusted — and what evidence is on record when it is followed.
Hallucination is unavoidable
Trustworthiness therefore comes from architecture, not from a better prompt: grounding, guardrails, abstention, and auditable evidence.
The lab
A physical lab at Luleå University of Technology, running the compute continuum in three always-on tiers.
Far edge
Reasoning at the asset
Handheld Raspberry Pi 5 devices, sensor probes, inspection cameras and mobile robots, running 1–7B quantized models directly on the machine.
Edge inference
Frontier models on site
On-prem GPU servers with 192 GB of GPU memory each, running frontier reasoning models locally. Data never leaves the site.
Cloud · HPC
Burst capacity on demand
Overflow to cloud and HPC when a job outgrows the site, and only then.
ColonyOS places each workload at the right tier.
Example applications
Building intelligence
3D building twin the agent navigates, queries, and acts on in the loop.
Wall thickness analysis
Inspection workflows over ultrasonic measurement data, with auditable reports.
Robot mower with onboard AI
A robot you talk to: the agent runs mowing missions, reads its own hardware logs, and explains itself. This demo — simulator, robot and UI — was generated and verified by AI agents in two days.
PhD projects
Three research directions, one question: what does it take to trust autonomous systems in the physical world?
Safe autonomy over live systems
When can an AI agent be allowed to act — and when must it hold back? Architectures where language models diagnose and decide while grounding, guardrails, and digital twins supply the restraint, so autonomy earns trust instead of assuming it.
Selected result5/5 correct abstention on honest load, zero false actions
The self-aware continuum
Orchestration across IoT, edge, and cloud where the application tells the infrastructure what matters right now. Scheduling, scaling, and offloading driven by runtime relevance instead of static thresholds — less cloud, better service, at industrial scale.
Selected result−24% cloud hours on production mining workloads
Trust without connectivity
Multi-stakeholder cyber-physical systems that keep operating through disconnection and still converge on one auditable truth: who was allowed to do what, when, and on whose authority. Decentralized authorization, tamper-evident history, deterministic reconciliation.
Selected resultDeterministic deny-wins replay, verified in Rust
Teaching
D7065E — Embedded Intelligence at the Edge (7.5 ECTS, Luleå University of Technology). Course notes, hands-on tutorials, labs, and a building-simulation server, all open: github.com/eislab-cps/D7065E.
Work with the lab
Commissioned research
Your problem, our researchers. Industrial PhD option, publishable results.
EU / Vinnova projects
Co-funded innovation at scale, with consortium partners across Europe.
CPS Lab partnership
Shared infrastructure, joint pilots, and a seat in the centre.
Innovation / MVP track
From idea to working prototype on a short horizon, via LTU Business.
Bring a real problem to the lab. Get in touch