BRENNEN

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

Scheduling

Scheduling decides which job runs on which machine, in what order and when, so that parts ship by their due dates and machines are not left standing idle.

6

Milestones achieved, from 1917 to 2008.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1917

    Work drawn against time

    The management engineer Henry L. Gantt had for some years used a chart on which the work for each machine was "laid out" along a scale of time. In 1917 General William Crozier, the US Army's Chief of Ordnance, hired him, first at the Frankford Arsenal, and Gantt turned it into a progress chart that compared work done with work promised; during 1918 it was used in US arsenals, in naval aircraft production and by the Emergency Fleet. Wallace Clark's 1922 book on the charts shows two kinds made for machine shops: a machine record chart showing each day which machines ran and why the others stood idle, and a load chart showing how much work was booked on each machine tool. Gantt never patented his charts and gave samples to anyone who asked, Clark wrote.

    archive.org

    Credit

    Henry L. Gantt; Wallace Clark

  2. 1953

    A rule for two machines

    Selmer M. Johnson of the RAND Corporation showed that when each of a batch of jobs must pass through one machine and then a second, a simple decision rule gives the order with the shortest total time, setups included. His paper says the rule lets one "optimally arrange twenty production items in about five minutes by visual inspection."

    rand.org

    Credit

    Selmer M. Johnson, RAND Corporation

  3. 1967

    System 24 stops short

    In September 1967 Theo Williamson of Molins, a London maker of cigarette machines, presented System 24: a group of computer-controlled machine tools fed with parts fixed to standard pallets, meant to run around the clock with staff on one shift and only maintenance staff on the other two. Molins put up a building for it at Deptford, and IBM planned a pilot at its plant in Rochester, Minnesota. By J. A. McGeough's 2025 biography of Williamson, new managers questioned the cost; while Williamson was in hospital in 1969 the work was effectively stopped, IBM withdrew, and in 1973 Williamson resigned. McGeough writes that Molins' patents still earned royalties on the flexible manufacturing systems, as such computer-run groups of machines came to be called, sold in later years.

    LessonThe idea outlived the project; Williamson concluded that new manufacturing technology needs "the backing of the big industrial players.".

    books.ed.ac.uk

    Credit

    D. T. N. (Theo) Williamson, Molins

  4. 1981

    Machines run through the night

    In October 1981 Yamazaki Mazak began an unmanned night shift at a flexible manufacturing plant at its Oguchi headquarters near Nagoya, Japan, by the company's timeline; IndustryWeek's 2004 profile of Mazak lists the same "lights out" system for 1981. The same profile reported that the machining cells and flexible manufacturing system at Mazak's plant in Florence, Kentucky, ran "24 hours a day with one shift unattended." FANUC, the Japanese maker of machine controls, says robot cells at its own plant reached 720 hours of continuous unmanned running in 2002.

    mazak.comindustryweek.comfanuc.co.jp

    Credit

    Yamazaki Mazak; FANUC

  5. 1989

    Ten jobs, ten machines, solved

    In the job-shop problem every job visits the machines in its own order, and the number of possible schedules grows explosively with size. A test case of ten jobs on ten machines, posed by J. F. Muth and G. L. Thompson in 1963, was solved with proof of the best answer "for the first time" by J. Carlier and E. Pinson in Management Science in February 1989, using branch and bound, a search that rules out whole families of schedules at once. In 1991 Applegate and Cook solved it in under 7 minutes on a Sun SPARCstation 1 computer.

    LessonA shop of only ten jobs and ten machines took 26 years to schedule provably best, and in 1991, Applegate and Cook wrote, "even modest sized instances" remained "computationally intractable.".

    api.crossref.orgapi.crossref.org

    Credit

    J. Carlier (Université de Technologie de Compiègne) and E. Pinson; Applegate and Cook

  6. 2008

    A common language for machines

    Each maker of machine controls had its own way of reporting what a machine was doing. In 2008 the Association for Manufacturing Technology (AMT), UC Berkeley's RAD Lab and the Georgia Institute of Technology launched MTConnect at the IMTS trade show in Chicago, an open, royalty-free protocol that lets software read a machine tool's state, so that a shop can time its parts, raise alarms and see whether machines are running as planned. Control Design reported that LNS Machine Tools of Cincinnati wrote its own adapter and showed live data six hours after starting. Germany's machine tool association VDW announced a rival, umati, nine years later, saying MTConnect could only read data and did not meet its members' needs.

    LessonTwo "common" languages confused buyers, by Cutting Tool Engineering's account in 2020, and their backers then agreed to work toward harmonizing them.

    controldesign.comctemag.com

    Credit

    AMT; UC Berkeley RAD Lab; Georgia Institute of Technology

Watch for

  1. Schedules that follow the floor

    Skody AI sells scheduling software that builds a plan aimed at on-time delivery from a shop's ERP data (enterprise resource planning, the business system that holds orders and routings) and re-plans as machines, people and materials change; users can click a recommendation to see the data behind it. Zach Yeager of Yeager Manufacturing, a job shop in Winter Park, Florida, told Modern Machine Shop that on-time delivery, which had topped out at 80% in good times and slipped to 70% during a rush of expedited orders, reached 85–90% within three months of adopting the software in February 2026.

    Achieved whena full year of on-time delivery records before and after adoption is published for this shop or others, checked by someone outside the shop and the vendor.

    mmsonline.com

    Claimed by

    Zach Yeager, Yeager Manufacturing, reported by Evan Doran in Modern Machine Shop, September 2026

Open

  1. 01

    Planned versus actual, in public

    An open set of anonymized records from at least five job shops, three months each: routings (the sequence of operations for each part), setup and run times, due dates, the schedule planned each morning, and what the machines actually did according to their own logs, including breakdowns, rush orders and late material. Add baseline re-plans from a constraint solver and from first-in, first-out, each scored on the number of late jobs.

    The brief
    Credit · unclaimed
  2. 02

    An open spindle-time census

    An independent count of how many hours a week CNC machines actually cut, from at least 50 shops in three or more countries, collected with open-source software and published as open data by machine type, shop size and shift.

    The brief
    Credit · unclaimed

Open 01 · the brief

Planned versus actual, in public

Credit · unclaimed

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

What to build

An open set of anonymized records from at least five job shops, three months each: routings (the sequence of operations for each part), setup and run times, due dates, the schedule planned each morning, and what the machines actually did according to their own logs, including breakdowns, rush orders and late material. Add baseline re-plans from a constraint solver and from first-in, first-out, each scored on the number of late jobs.

Why it is still open

Public test sets for scheduling are mostly made up. Reijnen and colleagues' open benchmark (2023, revised 2025) gathers classic job-shop problems and variants with setup times and arriving jobs; Deliktaş and colleagues' 2024 Data in Brief set has 43 instances, one of them from a real plant, with no record of what actually ran. Skody's on-time gain rests on one shop's statement. Searches in November 2026 found no public set that pairs a shop's plans with what its machines did.

arxiv.orgebi.ac.ukmmsonline.com

Built with what exists

Order and routing exports from each shop's ERP system; open MTConnect agents (mtconnect/cppagent, TrakHound's MTConnect.NET) for machine states; Google's OR-Tools CP-SAT solver for the baseline schedules; Python; Zenodo to publish.

  1. First step

    Ask one shop to export a month of jobs and due dates and to run an MTConnect agent on two machines, then write and publish the script that strips customer names and part numbers.

  2. Done looks like

    Five shops' three-month records published under an open license with a data dictionary and the baseline results, and at least one team outside the project reporting its own scheduler's results on them.

Open 02 · the brief

An open spindle-time census

Credit · unclaimed

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

What to build

An independent count of how many hours a week CNC machines actually cut, from at least 50 shops in three or more countries, collected with open-source software and published as open data by machine type, shop size and shift.

Why it is still open

The figure most often quoted, about 25% of the time, comes from MachineMetrics' opt-in customer data, presented by its co-founder Bill Bither and reported by Modern Machine Shop in 2020, against the 60% or more that shops assume before they measure; the underlying data are not public. Shops' own answers run far higher: in Modern Machine Shop's 2017 Top Shops benchmarking survey, with 356 responses, the leading fifth of shops reported spindle utilization of 72% and the others 60%. A GitHub search for "mtconnect" finds open data collectors but no shared dataset of machine use.

mmsonline.commmsonline.com

Built with what exists

The MTConnect agent (mtconnect/cppagent); Ladder99's fanuc-driver for FANUC controls; TrakHound's MTConnect.NET; a small computer such as a Raspberry Pi on each shop's network. Use only machines whose controls already have a network port, so that nobody has to open an electrical cabinet.

  1. First step

    Run the MTConnect agent on one shop's networked machines for a week and publish the anonymized state log with a one-page method.

  2. Done looks like

    50 shops and 500 machines logged for three months, with the data and method published under an open license, and the results set beside the vendors' published figures.