Edge AI
High-performance computing, called directly from your control program.
Edge AI gives the PLC runtime direct access to the GPU, NPU and CPU headroom on your device. AI models run as part of your control program, not as a separate system beside it. That covers language models, vision models or a model trained on your own process, all on the same device as the process and with no cloud in the loop. Edge AI is in beta, and we're opening early access to a small group of partners.
AI inside the controller, not bolted on beside it
One device, one runtime, one lifecycle.
Autonomy Edge
Deploys, monitors, rolls back
PLC runtime
AI function block
GPU / NPU
Control logic
I/O · fieldbus · sensors · cameras
Part of your program
The model is exposed as a function block in your IEC 61131-3 project, and its results arrive on block pins like any other input. There's no inference server to run, no protocol link to configure and no network hop. You get one download, one version and one audit trail for the logic and the model together.
Accelerator power, on control timing
The runtime reaches the device's GPU or NPU directly, so heavy models run at hardware speed. Inference runs separately from the scan cycle, and your program reads the latest completed result. A demanding model uses compute and power, but it doesn't cost you missed deadlines.
Your logic stays in charge
The model advises, and your program decides. Its output is treated like any other sensor reading and checked by logic your controls engineers can read and approve. If the model goes quiet, the machine falls back to a state you defined.
Nothing leaves the plant
Process data, images and model outputs stay on the device. The cloud sees deployment state and runtime health, never your tag values.
What you'll build with it
AI that answers in the same second the process needs it, on a model you chose.
What a model can do once it is this close
Ask the line a question
An operator asks why line 3 stopped, in their own language, and gets an answer built from live tags and the last hour of alarms.
Catch what a threshold can't
A model watches sensor streams for patterns nobody could write a setpoint for, running continuously without sending data off site.
Inspect the part
A vision model on the same device classifies what the camera sees and passes a verdict your logic can act on.
Collapse an alarm storm
Four hundred alarm rows reduced to the one root cause, before anyone has to read them.
What you deploy
Open-weight models
Language and vision models, quantized to fit your hardware, or a small model you fine-tuned on your own process.
Hardware-aware
Autonomy Edge knows each device's cores, memory and accelerators, so it only offers models the device can actually run.
Fleet deployment
Roll a model out to one device or to every device from the dashboard you already use, and roll it back the same way.
Runs offline
Once a model is deployed, the device needs no connection. If the link drops, inference keeps working, and so does the line.