BRENNEN

Industries / Mining and materials / Steel and metals / Machine shops

Programming machines

Programming turns a drawing or 3D model of a part into the list of moves, speeds and tool changes that a machine tool follows, today mostly through CAM (computer-aided manufacturing) software that writes the program.

6

Milestones achieved, from 1952 to 2003.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1952

    Numbers steer a mill

    Numerical control (NC) runs a machine tool from coded numbers, at first holes punched in paper tape, instead of a machinist turning handwheels. On 5 May 1952 John T. Parsons and his engineer Frank L. Stulen of the Parsons Corporation in Traverse City, Michigan, filed a patent for "automatically controlling machine tools, such as milling machines and the like, from media containing stored information"; it was granted in January 1958. The Massachusetts Institute of Technology (MIT) developed numerical control for the US Air Force in its Servomechanisms Laboratory, and MIT's 1959 television program on the work says its feasibility was proven in 1952. The National Inventors Hall of Fame, which inducted Parsons in 1993, ties his idea to the formulas used to design helicopter rotor blades.

    patents.google.cominfinite.mit.eduinvent.org

    Credit

    John T. Parsons and Frank L. Stulen, Parsons Corporation; MIT Servomechanisms Laboratory; US Air Force

  2. 1954

    Recorded motions lose out

    In 1954 General Electric (GE) developed one of the first numerical controls, a "record-playback" control that recorded a skilled operator's motions on magnetic tape and played them back to make duplicate parts, by American Machinist's account. GE called these its Mark Series; its next control, the Mark II Series, read its program from punched cards and then from punched paper tape. The historian David Noble argued in Forces of Production (Knopf, 1984) that automation was "shaped decisively by the U.S. military's preference for and lavish funding" of numerical control over a "middle-tech" record-playback option, "despite the absence of clear economic advantages", in the words of his memorial from the Society for the History of Technology.

    LessonIn Noble's account, the way of programming machines that won was the one the military paid to develop, not one shown to be cheaper, so who funds the work can settle a technical choice before shops do.

    americanmachinist.comhistoryoftechnology.org

    Credit

    General Electric; history by David Noble

  3. 1959

    A language for toolpaths

    APT, short for Automatically Programmed Tools, lets a programmer describe a part's surfaces and the cutter's path in word-like statements, and a computer works out the coordinates for the tape. Douglas T. Ross led the work as head of the computer applications group in MIT's Servomechanisms Laboratory, whose report "Automatically Programmed Tools, Volume I: General Description of the APT System" dates from about 1959, by the Computer History Museum's catalog. MIT's 1959 Science Reporter program, with Ross and the laboratory's director J. Francis Reintjes, says the idea took about two and a half months to prove and about two and a half years to develop, with 19 aircraft plants cooperating, and that one aircraft plant cut the time to program one part from 200 man-hours to 5. The programs ran on an IBM 704 computer at MIT.

    infinite.mit.educomputerhistory.org

    Credit

    Douglas T. Ross and J. Francis Reintjes, MIT Servomechanisms Laboratory, with 19 aircraft plants

  4. 1972

    Computers move into controls

    Fanuc's company history starts with a team set up in 1955 at Fuji Tsushinki Manufacturing, now Fujitsu, to develop numerical controls. Fanuc says the team developed the first NC and servo systems in Japan's private sector in 1956, shipped its first commercial control to Makino Milling Machine in 1958, finished what it calls the world's first commercial DNC (direct numerical control, in which one computer sends programs to several machines) in 1968, and introduced CNC in 1972. In computer numerical control (CNC), a small computer inside the control holds and runs the program. In 2024 NC machines made up 93.5% of the value of Japan's machine tool production, 843.1 billion yen of 901.3 billion yen, by Ministry of Economy, Trade and Industry (METI) figures for plants with 50 or more employees, published by the Japan Machine Tool Builders' Association (JMTBA).

    fanuc.co.jpjmtba.or.jp

    Credit

    Fanuc, which began as a team at Fujitsu

  5. 1979

    One code, many dialects

    The Electronic Industries Association completed EIA-274-D, its standard for numerical control programs, in February 1979, and the international version, ISO 6983, dates from 1982. Machinists call the language G-code; it mainly gives positions and feed rates for the machine's axes. In 2000 the US National Institute of Standards and Technology (NIST) wrote that implementations differ from machine to machine, and that "a program that runs on one machine probably will not run on one from a different maker". A 2005 report from VTT, Finland's state technical research center, said that because of dialects and vendor additions, part programs are not interchangeable between controls, so CAM software needs a separate postprocessor, a translator that rewrites its output for each make of control.

    LessonA standard that let each maker add its own commands did not make programs portable, so every make of machine still needs its own translator.

    tsapps.nist.govsarjaweb.vtt.fi

    Credit

    Electronic Industries Association; ISO

  6. 2003

    A richer standard stalls

    STEP-NC, published as the first parts of ISO 14649 in 2003, describes what is to be machined, the part's features, tools and process plan, instead of how to move the axes. It grew from European research projects that began in 1994 and then a program joined by the United States, Korea and Switzerland, according to the 2005 VTT report. In 2015 Mark Albert, editor emeritus of Modern Machine Shop, wrote that STEP-NC "seemed poised to transform the world of CNC machining, but the promised transformation didn't happen". Meanwhile CAM software got better toolpaths, postprocessors became more reliable and easier to get, and "conventional G-code programming held on as the prevalent format".

    LessonA better standard that needed new controls lost to better translation software that worked with the machines shops already owned.

    sarjaweb.vtt.fimmsonline.com

    Credit

    the ISO 14649 committees and the OPTIMAL, European STEP-NC and IMS STEP-NC research projects

Watch for

  1. AI drafts the CAM program

    CloudNC sells CAM Assist, a plug-in for Autodesk Fusion, Mastercam, Siemens NX and other CAM software that proposes machining strategies, tools, toolpaths and feeds and speeds for milled parts, which the programmer then reviews and edits. It covers 3-axis and 3+2-axis work (the part is tilted to a fixed angle, then cut with three moving axes) and does not yet cover lathes or mill-turn machines. CloudNC says the tool completes "up to 80%" of a CAM program in minutes and, in September 2025, that more than 1,000 machine shops use it. Its figures come from its own benchmarks and customer quotes, such as one shop's statement that a job that took an hour now takes 12 minutes; no independent test was found.

    Achieved whenan outside party times programmers with and without the tool on a published set of parts, and reports how much of each program needed changes and whether the cut parts passed inspection.

    cloudnc.comcloudnc.com

    Claimed by

    CloudNC, September 2025

  2. Toolpaths straight from 3D shapes

    Sangmin Hong, Dohee Cho, Mohsen Yavartanoo and Kyoung Mu Lee posted Geo2Gcode on 15 November 2026, a method that turns a 3D shape into a CNC program through a set of standard machining steps. In simulation on research CAD parts, they report roughing "comparable to an expert-tuned commercial CAM setup", with somewhat longer toolpaths. Their programs ran in a commercial G-code simulator, and one part was cut on a real mill. The paper links no code or data.

    Achieved whenthe code and test parts are released and another group cuts a set of parts with it and measures them against the models.

    arxiv.org

    Claimed by

    Hong and colleagues, November 2026

Open

  1. 01

    Scoring AI-written CNC programs

    A public test of CNC programs written by AI for real parts. Publish 100 part models with drawings, from flat brackets to 3+2-axis housings. Have each tool, from CAM plug-ins to general AI models, write the programs. Run every program in an open simulator to find gouges, collisions and missing features, cut a sample in machinable wax, and score each tool on how many programs were right without edits.

    The brief
    Credit · unclaimed
  2. 02

    Moving programs between controls

    A tested translator that converts a proven program from one control's dialect of G-code to another's, for example from Fanuc to Siemens or Heidenhain. It proves each translation by simulating both versions and comparing the shapes they cut.

    The brief
    Credit · unclaimed

Open 01 · the brief

Scoring AI-written CNC programs

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of CNC programs written by AI for real parts. Publish 100 part models with drawings, from flat brackets to 3+2-axis housings. Have each tool, from CAM plug-ins to general AI models, write the programs. Run every program in an open simulator to find gouges, collisions and missing features, cut a sample in machinable wax, and score each tool on how many programs were right without edits.

Why it is still open

Published tests are small or run by the makers. GLLM (Mohamed Abdelaal, Samuel Lokadjaja and Gilbert Engert, January 2025) scored a fine-tuned model on six shapes in the open-source CAMotics simulator. Šket and colleagues at the University of Maribor tested ChatGPT-3.5 and GPT-4o on one sample part in 2025 and report no success rate. Geo2Gcode (November 2026) reports simulation and one cut part, with no code. CloudNC cites no independent test. On GitHub, the repository LLM-Gcode-Eval, created in January 2026 to benchmark G-code from language models, held only a.gitignore file on 18 November 2026, and a December 2024 repository offers synthetic training programs for one Haas desktop mill. arXiv and GitHub searches in November 2026 found no shared set of parts with scored programs.

arxiv.orgapem-journal.orggithub.comgithub.com

Built with what exists

CAMotics, the open-source CNC simulator GLLM used; Python to compare each simulated part with its model; open-weights AI models run on a shop's own computer, such as Mistral Large 4, so that donated drawings stay private (see the Note "A trillion open parameters"); a desktop or school CNC mill cutting machinable wax; calipers and a height gauge.

  1. First step

    Collect 20 part models with drawings that a school or makerspace can share, and write the scoring sheet: runs in simulation, matches the model within 0.1 mm, no collision.

  2. Done looks like

    100 parts, every program and every score published under an open license; at least three tools tested; 20 parts per tool cut in wax and measured against the drawing; a table showing which programs ran right the first time, which needed edits and which would have crashed.

Open 02 · the brief

Moving programs between controls

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A tested translator that converts a proven program from one control's dialect of G-code to another's, for example from Fanuc to Siemens or Heidenhain. It proves each translation by simulating both versions and comparing the shapes they cut.

Why it is still open

NIST in 2000 and VTT in 2005 both said programs do not move between makes of control, and Modern Machine Shop wrote in 2015 that G-code programs "often became useless" when the original machine was replaced. Even one maker's programs do not always move: HEIDENHAIN's 2018 manual for its TNC 640 control, which can be programmed in ISO format as well as in HEIDENHAIN's own language, says programs written on its earlier controls "may not always run" on it. A GitHub project started in March 2026 translates Fanuc programs to Siemens 840D with a language model and checks the translated toolpath against the original in simulation, but it publishes no pass rate on a set of real shop programs. GitHub searches for "gcode dialect" and "heidenhain fanuc" in November 2026 found parsers, cheat sheets and postprocessors that write new programs from CAM, but no tested translator of existing programs.

mmsonline.comgithub.comtsapps.nist.govcontent.heidenhain.de

Built with what exists

An open G-code parser in Python; CAMotics or another simulator to compare the material each version removes; a language model for commands with no direct match, such as canned cycles (built-in routines for drilling or pocketing) and macros; old programs donated by shops with part names removed.

  1. First step

    Ask three shops for 20 retired programs each, with the matching part models, and write the comparison script for one pair of dialects.

  2. Done looks like

    200 real programs from at least three shops translated into two other dialects; at least 95% match the original's simulated cut within 0.01 mm; 10 translated programs cut in wax on machines with the target controls and measured; every failure listed with its cause.