Industries / Mining and materials / Steel and metals / Machine shops
Fixtures and setup
Setup is the work done before the first cut on a new job: clamping the part in a vise, chuck or fixture, loading the tools, and telling the machine where the part and the tools are.
4
Milestones achieved, from 1896 to 2015.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1896
Magnets hold the work
A magnetic chuck holds iron and steel parts flat on a magnetized plate, with no clamps in the way of the tool. Oakley S. Walker started the O.S. Walker Company in 1896 with the magnetic chuck as its main product; the company says he invented it. Industrial Magnetics, which bought Walker Magnetics in 2020, calls Walker "the inventor of the electromagnetic chuck" and his company the oldest industrial magnet maker in North America.
Credit
Oakley S. Walker
1977
A probe finds the part
A CNC machine has to know where the part and each tool sit before it can cut. David McMurtry invented the touch-trigger probe in 1972 for coordinate measuring machines, and Renishaw, which began selling his probe for those machines in 1973, says its first probe made for machine tools followed in 1977, "enabling automated setting and inspection" on CNC milling machines. The probe is held in the spindle like a tool and signals the instant it touches the part, so the machine can record where the part is. The Manufacturer reported in 2023 that this work now covers setting both tools and workpieces on the machine.
Credit
David McMurtry; Renishaw
1980s
Pallets that return to zero
A zero-point clamping system locks a pallet or fixture plate onto studs in the same position every time, so a job set up away from the machine can be dropped in without being measured again. EROWA, founded by Hans Hediger in a garage in Reinach, Switzerland, in 1970, says it developed its first ITS zero-point clamping system in the 1980s and its first robots in 1987. The method is now sold as a standard part of robot cells: FANUC America's mill cell has a robot swap fixture plates on zero-point studs, with a changeover "typically under one minute" by FANUC's account.
Credit
EROWA and its founder Hans Hediger
2015
Robots change the vise jaws
A robot that loads parts usually needs a person to change the workholding between jobs. In March 2015 Modern Machine Shop reported on VersaBuilt's system, in which the vise jaws that hold a part in the machining center also serve as the robot's grippers, and showed it running three operations on one part unattended on a vertical machining center, with the robot switching jaws between operations.
Credit
VersaBuilt
Watch for
Probing in place of fixtures
GTR Composites Group, a UK maker of composite race-car chassis, says aligning each chassis on its machine took up to 15 hours, because parts warp when they come out of the mold. It says that since adopting NC-PartLocator, software from Metrology Software Products (MSP) that probes the part on a five-axis Zimmermann FZU400 machine and computes a corrected alignment, it saves 10 hours per chassis and no longer builds a fixture from scratch for each one.
Achieved whenan outside party measures setup time and finished-part accuracy on a run of parts with and without the software.
Claimed by
GTR Composites Group and MSP, in a case study published by Metrology News, November 2026
Open
01
Robots that learn vise loading
A robot that learns to load raw blocks into a machine vise and unload finished parts from a few human demonstrations, instead of from a program written for each part, with every loaded part checked by the machine's own probe to see whether it is seated.
The briefCredit · unclaimed02
Workholding plans from the model
A program that reads a part's 3D model and proposes how to hold it for each setup: vise, soft jaws (jaws machined to the part's shape) or a fixture plate, which faces are clamped, and how many setups the part needs. Test it against plans written by machinists, then cut some parts by its plans.
The briefCredit · unclaimed
Open 01 · the brief
Robots that learn vise loading
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A robot that learns to load raw blocks into a machine vise and unload finished parts from a few human demonstrations, instead of from a program written for each part, with every loaded part checked by the machine's own probe to see whether it is seated.
Why it is still open
The robot cells in use are programmed for each job: VersaBuilt's jaw-changing system and FANUC's zero-point cell move parts along taught or computed paths. Intrinsic released an open machine-tending application (OMTS) under the Apache 2.0 license in July 2026; it runs a programmed behavior tree and finds parts with NVIDIA's FoundationPose pose estimator, and its README gives no success counts from a real machine. A search of arXiv for "machine tending" found work on assigning tasks across several robots and machines (2024, 2026) and on human-supervised mobile robots (2020), but no published tally of a robot that learned vise loading from demonstrations and was checked for seating on parts it had not seen. A GitHub search for "machine tending" found small projects, such as one documenting a Kinova Gen3 arm tending a Tormach 770M mill, with no published success rates.
Built with what exists
A benchtop CNC mill with a door interlock, run so the robot works only with the spindle stopped and the door open; a vise with a powered or pneumatic clamp; a low-cost SO-101 arm with Hugging Face's LeRobot software for a first try (see the Note "No download"), or a collaborative arm such as a Universal Robots UR5e; NVIDIA's GR00T 2 robot model, which by NVIDIA's own test averages 95% success after five demonstrations (see the Note "Five demonstrations per task"), or Runway's open Praxis-1 (see the Note "Open robot model weights"); a spindle probe to check seating; Intrinsic's OMTS as a programmed baseline.
First step
Lock out the spindle of a benchtop mill and record 20 demonstrations of loading one aluminum block into its vise.
Done looks like
200 load-and-unload cycles over ten part shapes, five of them never shown in training, with every loaded part probed. The published table gives the share seated within 0.05 mm of the reference position for the learned robot and for the programmed baseline, with video of every failure.
Open 02 · the brief
Workholding plans from the model
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A program that reads a part's 3D model and proposes how to hold it for each setup: vise, soft jaws (jaws machined to the part's shape) or a fixture plate, which faces are clamped, and how many setups the part needs. Test it against plans written by machinists, then cut some parts by its plans.
Why it is still open
Most tools on sale leave the choice to a person, and the one that claims to make it publishes no test. Toolpath, which sells AI CAM software that runs inside Autodesk Fusion, says it "proposes how to hold and orient the part across setups" and can export the workholding to Fusion, but its product page gives no accuracy figures. CloudNC made its Soft Jaw Designer free in July 2025; it cuts the jaw shape after the user places two jaw blanks around the part, and CloudNC says modeling soft jaws by hand takes machinists 30 to 60 minutes per setup. MillBill, an open agent posted on arXiv in November 2026 by Savrasov and colleagues, plans setups for three-axis milling by what the cutters can reach and scores the results in simulation, but it does not choose clamps or check that the part stays put. An arXiv search for fixture design and setup planning with learning or language models found a 2022 method for laying out fixtures and no benchmark that compares AI workholding plans with machinists'.
Built with what exists
An open CAD kernel through CadQuery or pythonOCC; MillBill's open code for setup orientations; a vision-language model to read the part's drawing; Google's Gemini Decision 1, which returns a probability for each answer on a fixed list, to pick among holding methods (see the Note "A thousand questions, one pass"); the 11 free test models with full tolerances from NIST's model-based enterprise project.
First step
Ask three machinists to write setup plans for the 11 NIST test models.
Done looks like
On 60 parts, each with plans from three machinists, the program's plan matches at least one machinist's on the number of setups and the clamped faces for 80% of parts. Ten parts are then cut in aluminum by the program's plans, and none moves in its holding, checked by measuring the finished features against the drawing.