Use cases · Actuation · Industrial automation
Say it. The plant does it.
The PLC proves it.
Type a plant action in plain language in the Senua AI console: “start pump one”, “power up fan one”, “stop all fans”. Senua drafts it into a visual control sequence, runs it against a real PLC, and every write is proven by reading the device back : on-prem, offline, no vendor SDK. Commanding a machine is the same causal-state engine that flies the drone, run forward, with a wire-protocol codec where the radio used to be.
The scenario
The shift supervisor who doesn’t open the HMI.
Priya runs the night shift at a food-processing plant: three lines, hundreds of motors, valves and sensors, two people on. Every routine action means finding the right HMI screen among dozens, or radioing the one technician who knows where the tag lives. The plant is air-gapped; cloud AI tools are banned outright, and she would not trust a system that can hallucinate a register write to touch her lines anyway.
She wants to state the action and have it executed and verified against the controllers the plant already runs: on-site, with every write proven by the device’s own readback, and a refusal when the system isn’t sure. Routine operations should stop depending on HMI archaeology and radio calls.
The console, live, beside the plant
One screen, split in two: the Automation editor on the left, a real PLC runtime on the right: driven by the exact same writes, in real time.
1 · Compose in plain language
The Automation editor’s composer turns an intent: “start pump one”, “open up valve one”: into steps on a canvas (turn coil on, set register, read coil) against a controller target. The sequence is generated by the learned specialist, the same engine as the drone mission: no keyword parser, no template library to maintain.
2 · Run it on a real PLC
Press Run and the pure-Senua Modbus codec executes the sequence against OpenPLC v3 : a genuine PLC software stack, scan cycle running, no simulator shims. The engine speaks the wire protocol byte-by-byte; there is no vendor library and no runtime dependency.
3 · The machine is the truth
Every write is followed by a readback through the same protocol; the run passes only if the device state matches the intent. On the right, the PLC’s own monitoring page shows the plant points flip on in real time: the audit trail is the machine itself, not a judge model.
Why it’s different
One engine, many bodies. The wire protocol, not the SDK.
The same substrate that flies the drone.
The specialist that generates these PLC sequences is byte-for-byte the same one that drafts drone missions. Adding the industrial target added a wire-protocol codec and a corpus: zero changes to the cognition engine. A new machine is a data operation, not a new AI project.
It runs where industrial systems live.
The engine speaks Modbus TCP directly, frames built in the substrate: no SDK, no cloud account, no egress. It deploys offline, air-gapped, on commodity CPU, exactly where a security review will actually allow it.
Verified by the device, not the model.
The pass condition is not “the output looked right”: it is the PLC’s own state agreeing with the intent, read back through the same protocol, per write, including multi-device intents.
Honest by construction.
An intent the specialist has not learned yields an explicit abstention, never a guessed write to a live machine. On a plant floor, the confident wrong answer is the catastrophic one.
Scope & claims discipline
Measured before it was shown.
Generation validity was measured before the demo was recorded: 23 of 24 held-out corpus-regime intents were generated exactly and executed green against an independent controller oracle, every write verified by readback. The scope is the validated target (Modbus TCP / OpenPLC runtime) and corpus-regime intents: capability learned through use, the same model as the mission specialist. Novel-phrasing generalization is a measured open finding with a named next step, not a claim; and there is no real-time motion-control claim: the PLC’s scan cycle retains the safety role, exactly as the autopilot does in the drone story.
Say it to your own controllers.
Start on your own test cell: your tags, your corpus, verified readback on a non-production line, entirely on-site. The engine that watches the machine’s health is the engine that acts on it: one mind, both directions of the same wire.