Industries / Industrial equipment and services / IoT and industrial automation / Industrial controls and PLCs
Programmable controllers
A programmable controller is the computer, in its own rugged box or as software on an ordinary one, that reads a machine's inputs and sets its outputs many times a second by following a program.
5
Milestones achieved, from 1996 to 2024.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1996
Control becomes Windows software
Beckhoff Automation sold its first PC-based machine controller in 1986, and in 1996 it launched TwinCAT, software with real-time extensions that turns a Windows industrial PC into a PLC and motion controller programmed under IEC 61131-3. Beckhoff now quotes cycle times of 2 ms, 250 µs or 50 µs and more than 1,000 motor axes from one PC, its own figures. The market research firm IoT Analytics put Beckhoff's sales at more than 30% of those of Rockwell Automation's factory automation division in 2019, up from under 5% in 2000.
Credit
Hans Beckhoff and Beckhoff Automation
1998
The PLC declared dead
In 1998 a trade magazine asked whether desktop PCs would run the machines on the factory floor, and the people it quoted said the PLC was finished, Nikki Gonzales wrote in Packaging World in August 2026. The Windows-based controllers of the 1990s, among them ASAP, Think and Do, Steeplechase Software and Wonderware, did not win lasting use; IoT Analytics says "blue screens of death" raised doubts about their reliability, and it estimated that software PLCs held 3.5% of the controller market in 2019. Packaging World also cites an Automation World survey from around 2008 that found programmable automation controllers (PACs), the next challenger, in about 5% of the applications that could use them.
LessonPlants buy a controller to run and be supported for 20 to 50 years, so a cheaper computer that might crash did not replace a box that does not.
Credit
the PC-control makers of the 1990s, as reported by IoT Analytics and Packaging World
2018
An open PLC for research
Thiago Alves and Thomas Morris published OpenPLC in the journal Computers & Security in September 2018: an open-source PLC runtime and editor that follows IEC 61131-3, written so that researchers could study controller security on a PLC whose code they could read. Crossref counted 78 works citing the paper by November 2026, and the GitHub repository for version 3 of the runtime had about 1,600 stars and 617 forks before it was archived in April 2026. Autonomy, the company that now develops it, reports more than 1 million downloads, its own figure.
Credit
Thiago Alves and Thomas Morris
2018
Floods slow the scan
Matthias Niedermaier and colleagues at Hochschule Augsburg and Freie Universität Berlin flooded 16 controllers from six makers, among them Siemens, WAGO, Phoenix Contact, ABB, Crouzet and Schneider Electric, with network traffic while timing the electrical outputs each one drove. All but one were affected, some slowing and some stopping their outputs, and in three cases the traffic shut down the electrical side entirely, halting any process the controller ran. When the US agency ICS-CERT later published the makers' responses, only Schneider Electric had released patches; ABB said its device had been left in its default configuration, and Siemens said the report showed no vulnerability in its products.
LessonTraffic aimed at a controller's network port can disturb the machine it runs, and the makers did not agree that this was a flaw.
Credit
Matthias Niedermaier, Jan-Ole Malchow, Florian Fischer, Daniel Marzin, Dominik Merli, Volker Roth and Alexander von Bodisco
2024
Real-time Linux joins mainline
Linux 6.12, released on 17 November 2024, included PREEMPT_RT, the real-time patch set that makes as much kernel code as possible preemptible (able to be interrupted by a more urgent task) so that response times stay short and predictable. It had been developed outside the main kernel for 20 years. Software PLCs run on such kernels: Intel's Edge Controls for Industrial software, which offers PREEMPT_RT kernels, includes a CODESYS benchmark that runs a control task every 250 µs on processor cores set aside for it.
Credit
the Linux kernel's PREEMPT_RT developers
Watch for
A controller sold as post-PLC
Vention, a Montreal automation company, says its MachineMotion AI controller, built on NVIDIA's Jetson computer, runs robots, conveyors and vision models from one box and drives up to 30 servo motors over EtherCAT, marking "the post-PLC (programmable logic controller) era." In August 2026 Vention told Packaging World that the phrase does not yet hold for certified machine safety, that full commercial shipping began only in late 2025, and that it could not name a running customer or share a metric.
Achieved whena named customer or an outside party reports production machines run by it, with uptime figures.
Claimed by
Vention, September 2024, restated in August 2026
Open
01
AI beside the control loop
An outside test that runs an AI model, such as an image check or an anomaly detector, on the same industrial computer as a software PLC, and measures how much the PLC's cycle timing wanders (its jitter) and how many cycles it misses, with and without the model running.
The briefCredit · unclaimed02
One program, every maker's timing
A public benchmark: a fixed set of control programs (on-off logic, integer arithmetic, floating-point control loops, text handling) ported to controllers from many makers, with each controller's scan time, the time for one pass through the program, measured from outside the controller.
The briefCredit · unclaimed
Open 01 · the brief
AI beside the control loop
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An outside test that runs an AI model, such as an image check or an anomaly detector, on the same industrial computer as a software PLC, and measures how much the PLC's cycle timing wanders (its jitter) and how many cycles it misses, with and without the model running.
Why it is still open
Vention and CODESYS both describe keeping motion control on isolated processor cores while AI inference runs separately on a graphics processor, but Vention could share no metric. Intel's published CODESYS benchmark measures jitter on isolated cores with no AI load, and Beckhoff's documentation reports the latency of the AI model itself, not its effect on the control cycle.
Built with what exists
Intel's CODESYS benchmark program and a PREEMPT_RT Linux kernel; the OpenPLC runtime; an NVIDIA Jetson board or a PC with a graphics card; ONNX Runtime to run open image models; a logic analyzer timing an output that the PLC toggles every cycle.
First step
Run Intel's CODESYS benchmark on an Intel PC with a PREEMPT_RT kernel for 24 hours, then repeat with an image model running on the same machine and compare the two jitter records.
Done looks like
For at least three platforms, 72-hour runs at a cycle of 1 ms or shorter, with and without an AI model running, with maximum jitter and missed cycles published along with the scripts.
Open 02 · the brief
One program, every maker's timing
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public benchmark: a fixed set of control programs (on-off logic, integer arithmetic, floating-point control loops, text handling) ported to controllers from many makers, with each controller's scan time, the time for one pass through the program, measured from outside the controller.
Why it is still open
The open benchmarks found cover one maker or one runtime: Intel's for the CODESYS runtime on Intel processors, and Johannes Kutning's set of benchmark programs, written in STL for Siemens' S7-300 and S7-1500 only. A retailer's knowledge base compares two workloads on controllers from four makers, but timed by each controller's own timer, with array sizes cut to fit smaller controllers and one result scaled up tenfold.
Built with what exists
IEC 61131-3 Structured Text, the text language most makers' tools accept; AI coding agents to port each program into each maker's dialect; second-hand controllers from several makers; OpenPLC and CODESYS-based controllers as low-cost entries; a logic analyzer to time an output toggled once per scan.
First step
Write the on-off logic workload in Structured Text, run it on OpenPLC and on one CODESYS-based controller, and time an output toggle with a logic analyzer.
Done looks like
Five workloads run on at least ten controllers from at least five makers, with the programs, raw timings and jitter histograms published, and each maker invited to reply.