Industries / Industrial equipment and services / IoT and industrial automation / Industrial controls and PLCs
Programming
Programming is the writing, testing and managing of the programs that tell a controller what to do.
7
Milestones achieved, from 1977 to 2024.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1977
Steps and transitions
In 1975 a working group on logic systems at AFCET, the French automatic control association, set out to standardise how engineers describe sequential control. By 1977 it had defined GRAFCET, a chart in which a machine moves from one step to the next when the condition on each transition is met, and in 1988 the IEC made it an international standard, now IEC 60848, which calls it a language "for sequential function charts."
Credit
Michel Blanchard and AFCET's "Logic Systems" working group of industrialists and academics
1994
A tool for any maker
In July 1994 Dieter Hess and Manfred Werner founded 3S-Smart Software Solutions in Kempten, Germany, and launched CoDeSys, a programming system for the IEC 61131-3 languages that controller makers license instead of writing their own. The company, renamed CODESYS GmbH in 2020, counts 2 million new licences a year and hundreds of makers of controllers and automation parts among its customers.
Credit
Dieter Hess and Manfred Werner, 3S-Smart Software Solutions
2010
Controllers and screens together
On 23 November 2010 Siemens introduced the TIA Portal (Totally Integrated Automation Portal), one program in which its controllers and operator panels are configured, with its drives to follow. Tags, blocks and messages are entered once and shared, where before each device type had its own software package.
Credit
Siemens Industry Automation Division
2018
Unit tests for PLCs
In December 2018 Jakob Sagatowski released TcUnit, a free framework under the MIT licence for unit testing, the practice of checking each small piece of a program automatically against expected results, in Beckhoff's TwinCAT 3. On 19 November 2026 it had 426 stars and 87 forks on GitHub, by GitHub's count, and its author described it as "stable but resting."
Credit
Jakob Sagatowski
2019
A format for moving programs
PLCopen, a maker-independent association founded in 1992 and based in the Netherlands, wrote an XML file format for IEC 61131-3 programs that records graphical layout as well as text, and in 2019 the IEC adopted it as IEC 61131-10. PLCopen says users wanted to exchange programs between tools after the 1993 language standard, "although this was not the intent of the standard itself."
LessonA shared programming language did not make programs portable between makers' tools; that took a shared file format as well.
Credit
PLCopen's TC6 working group
2023
A chatbot writes control code
Three researchers at ABB Research in Ladenburg, Germany, wrote 100 programming requests in 10 categories and put them to ChatGPT running GPT-4. It "generated syntactically correct IEC 61131-3 Structured Text code in many cases," and they offered the request set as the basis for a benchmark.
Credit
Heiko Koziolek, Sten Gruener and Virendra Ashiwal, ABB
2024
Ladder logic resists models
Yimin Zhang and Mario de Sousa found that a language model given a few examples could draw a correct sequential function chart for simple requirements, but that "generating a Ladder Diagram (LD) automatically remains a challenge even for very simple use cases." They note that the graphical languages still have the most users.
LessonProgress by AI on text programs does not carry over by itself to the graphical language most PLC programmers use.
Credit
Yimin Zhang and Mario de Sousa, University of Porto
Watch for
Copilots in makers' tools
Siemens says its Eigen Engineering Agent, an AI assistant connected to TIA Portal, writes and adapts code in SCL (Siemens' Structured Text) and ladder, generates test logic and fixes compile errors, with "2–5x faster execution" and "up to 50% higher efficiency." Schneider Electric's Philippe Rambach wrote that its PLC Code Generation Copilot in EcoStruxure Automation Expert "reduces development time by 30–50%." Neither says how the figures were measured.
Achieved whena customer or outside lab publishes engineering time and error counts for the same work done with and without the assistant.
Claimed by
Siemens, product page read November 2026; Schneider Electric, November 2025
Control logic from narratives
ABB researchers say their Spec2Control workflow turns a written control narrative into graphical control logic for a distributed control system. On an open set of 10 narratives and 65 test cases it produced 98.6% of the correct connections by itself and would save 94–96% of human labour. ABB says it is building the workflow into its engineering tools and has released an open-source version "for independent validation."
Achieved whena group outside ABB runs the open version on its own plant narratives and publishes comparable figures.
Claimed by
Heiko Koziolek and five co-authors, ABB, October 2025
An agent held to tests
Researchers at Midea's AI Research Center describe SemaPLC, an agent harness that may only finish a PLC programming task once compile and live-runtime checks pass. Across 182 tasks and seven models its mean verified-pass rate was 72.6%, and in KUKA's iiQworks PLC tool, they write, it "substantially improves PLC development efficiency by 45%."
Achieved whenKUKA's users or an outside group publish the measurement behind the 45% figure.
Claimed by
Yanlun Tu and 13 co-authors, Midea AIRC, August 2026
Open
01
A public ladder test set
A public set of ladder-logic tasks, each with a plain-language description, a reference ladder program in PLCopen XML and input sequences with expected outputs, plus a harness that loads any AI model's answer into an open-source PLC runtime, runs the sequences and scores it.
The briefCredit · unclaimed02
Tests for a whole library
An AI agent that writes unit tests for each function block of OSCAT BASIC, an open-source library of IEC 61131-3 blocks, runs them, and grades them by mutation score, the share of deliberately planted bugs the tests catch.
The briefCredit · unclaimed
Open 01 · the brief
A public ladder test set
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of ladder-logic tasks, each with a plain-language description, a reference ladder program in PLCopen XML and input sequences with expected outputs, plus a harness that loads any AI model's answer into an open-source PLC runtime, runs the sequences and scores it.
Why it is still open
The authors of an ACL 2025 paper on ladder generation, from POSTECH and Hyundai Mobis, built their own data "due to the lack of publicly available datasets," could not release it for confidentiality reasons, and left "a publicly available VPL-based benchmark" as future work; their best method reproduced the reference program exactly in at most about two-thirds of cases. Zhang and de Sousa's September 2025 ladder datasets pair each program with a sequential function chart for converting one into the other, and a July 2026 suite of 50 ladder and Structured Text programs from the University of Manchester serves formal verification tools; neither sets tasks in plain language.
Built with what exists
OpenPLC Editor and Beremiz, open-source editors that export ladder programs as PLCopen XML; the OpenPLC runtime to execute them; ESBMC-PLC, the Manchester group's open verifier, to check properties; AI models such as Claude, Gemini and Qwen.
First step
Write 20 tasks from textbook exercises such as a conveyor, a traffic light and a tank fill, draw the reference programs in OpenPLC Editor, and score one model.
Done looks like
300 tasks covering interlocks, sequences, timers, counters and alarms, each reference program passing its own tests, and at least five models scored on the share of tasks whose program passes every test sequence, all released under an open licence.
Open 02 · the brief
Tests for a whole library
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An AI agent that writes unit tests for each function block of OSCAT BASIC, an open-source library of IEC 61131-3 blocks, runs them, and grades them by mutation score, the share of deliberately planted bugs the tests catch.
Why it is still open
In ABB's 2024 study a model wrote tests for 10 OSCAT blocks with high statement coverage on simpler ones, but the tests had "erroneous assertions in many cases, which still require manual adaption." STMutants, which its authors call the first public mutation benchmark for Structured Text, covers 11 programs and 108 planted bugs (June 2026).
Built with what exists
the OSCAT BASIC library; the STMutants mutation operators; TcUnit on TwinCAT 3, or the OpenPLC runtime; AI coding agents such as Claude Code.
First step
Run an agent on the 11 STMutants programs and compare its results with the paper's.
Done looks like
Tests for at least 100 OSCAT blocks, each published with its mutation score, a median score of at least 80%, and a count of the assertions a person had to correct.