Industries / Technology / Electronics and semiconductors / Semiconductor equipment
Wafer handling
Wafer handling covers the sealed boxes, robots, ceiling vehicles and stores that move wafers between and inside machines, and the message standards that let a factory's computers run those machines.
7
Milestones achieved, from 1984 to 2022.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1984
Wafers ride in sealed pods
SMIF (standard mechanical interface) is a sealed box, or pod, that carries a cassette of wafers and keeps clean air only around the wafers and inside the tool, instead of in the whole room. It was developed at Hewlett-Packard, which did not see it as core business and released it to Mihir Parikh, who founded Asyst to sell it; Asyst was incorporated on 31 May 1984.
Lessona box alone does not keep wafers clean; the same account says IBM's earlier portable enclosures "proved to be no cleaner than the factory air, usually dirtier".
Credit
Mihir Parikh and Tony Bonora, who received the 1999 SEMI Award for it; Hewlett-Packard; Asyst
1987
One robot, many chambers
A cluster tool places several single-wafer process chambers around a central chamber, where a robot moves each wafer in turn from a load lock (a small chamber pumped to vacuum so wafers enter without letting in air), so a wafer can get several steps without meeting air. Applied Materials' Precision 5000 of 1987 had four chambers; the Chip History Center calls it "the first successful system" of its kind to be placed into manufacturing, while the Semiconductor History Museum of Japan dates earlier multi-chamber machines to Nichiden Varian (1978) and Drytek (1984). In 1990 Applied's Endura 5500 joined two transfer chambers, so high-vacuum sputtering and lower-pressure steps could run in one machine.
Lessonkeeping the wafer inside the machine between steps cut contamination and waiting time, the museum says.
Credit
Dan Maydan, David N. K. Wang and Sasson Somekh, Applied Materials; Drytek
1992
GEM gives tools common behaviour
By the late 1980s tools talked to the factory's host computer using SEMI's SECS messages, sent over a serial cable as set out in the SEMI E4 standard. In 1992 SEMI published the first version of E30, the Generic Equipment Model (GEM), which adds a common set of messages and behaviours, so the factory's software can run and collect data from any maker's tool in the same way. Cimetrix, a supplier of this software, says these standards are now required on nearly every 300 mm tool.
Credit
SEMI standards committees
2001
Pods too heavy for hands
For 300 mm wafers the carrier became the FOUP (front-opening unified pod), a sealed box whose front door is opened by the tool's load port, its docking station. Asyst engineers wrote in 2001 that 300 mm carriers have 2.5 times the weight and volume of 200 mm ones, and Entegris lists its FOUP at 7.3 kg with wafers. Overhead hoist transport, small vehicles on ceiling rails that lower a pod straight onto a load port, took over the carrying.
Lessoneach overhead drop takes 20 to 25 seconds and blocks the rail meanwhile, and fixing the carrier at 25 wafers later made small batches hard to run.
sst.semiconductor-digest.comsst.semiconductor-digest.comentegris.com
Credit
Tony Bonora and David Feindel, Asyst (Solid State Technology, 2001); Entegris (FOUP data)
2009
Asyst goes bankrupt
Asyst, maker of SMIF pods, load ports and automated material handling systems (AMHS), said in its annual report for 2007 that it believed it led the AMHS market. It filed for Chapter 11 bankruptcy on 20 April 2009, saying demand for chipmaking equipment had "declined dramatically" in the recession, sold its businesses to Crossing Automation, the PEER Group and Murata Machinery, and stopped all business on 1 September 2009.
Lessona supplier that depends on new fabs being built can run out of cash when building stops, even while its products live on under new owners.
Credit
Asyst Technologies
2018
A new tool spreads malware
On the evening of 3 August 2018 a variant of the WannaCry virus infected fab tools and automated material handling systems at TSMC, making tools crash or reboot over and over. TSMC said a new tool had been connected to its network during software installation without first being isolated and confirmed free of viruses; about 80% of affected tools were working again two days later, and TSMC expected the incident to cost no more than 2% of its third-quarter revenue.
Lessonevery machine that joins a factory network is a way in, so it has to be isolated and checked first.
Credit
TSMC (disclosure)
2022
A security floor for tools
SEMI published what it called its first cybersecurity standards: E187 sets minimum security for fab equipment in four areas (operating systems, network security, endpoint protection and security monitoring), and E188 requires suppliers to scan equipment, and the devices used to maintain it, for malware before shipment. In 2023 TSMC made E187 part of its rules for buying tools, and Taiwanese reporting in December 2025 said vendors supplying TSMC must meet it and pass third-party checks.
Credit
a task force led by Leon Chang (TSMC) and Ares Cho (ITRI) for E187; Ryan Bond (Intel) and Richard Howard (Cimetrix) for E188
Watch for
Humanoids load wafer carriers
STMicroelectronics' head of manufacturing, Thomas Morgenstern, showed a video of a humanoid robot placing a wafer carrier into a machine, said "This is the first one we have", and said the company expects "numbers beyond one hundred humanoids" in its plants in the next couple of years; separately, GlobalFoundries will begin testing Boston Dynamics' Atlas humanoid in early 2027, with tasks, numbers and sites not disclosed. Note: "A real production line" covers a humanoid trial on car-parts lines.
Achieved whenan outside party reports humanoids routinely moving carriers in a production fab, with counts or uptime figures.
Claimed by
STMicroelectronics, March 2026; GlobalFoundries and Boston Dynamics, November 2026
AI reads tool interface manuals
Ergotech says an agentic AI system, one that works through a task in a chain of smaller steps, can read and interpret a tool's SECS/GEM specification, the document an engineer now reads by hand to connect a new tool to the factory, and that its approach "reduces the likelihood of errors and hallucinations". It gives no accuracy figures.
Achieved whena chipmaker or toolmaker reports using it on real manuals, or an independent test publishes its accuracy.
Claimed by
Jim Redman, Ergotech, July 2025
Open
01
Dummy-wafer transfer by hobby arm
Train a low-cost robot arm to move dummy wafers from slot to slot between two open cassettes, and publish the dataset, the trained policy and a filmed tally.
The briefCredit · unclaimed02
Overhead transport scoreboard
A frozen public benchmark for dispatching and routing overhead hoist vehicles on the SMAT2022 fab layout, run in the open U-FAST simulator, with a scoring script and a leaderboard outsiders can submit to.
The briefCredit · unclaimed03
Test AI on GEM manuals
A public set of interface manuals for imaginary tools, each with a hidden answer key listing the tool's status variables, collection events, alarms and remote commands with their numeric IDs, and a script that scores how well a model extracts them into a fixed JSON format. This scores reading documents, not whether a tool's software behaves correctly.
The briefCredit · unclaimed
Open 01 · the brief
Dummy-wafer transfer by hobby arm
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Train a low-cost robot arm to move dummy wafers from slot to slot between two open cassettes, and publish the dataset, the trained policy and a filmed tally.
Why it is still open
A Hugging Face search for "wafer" returns 29 datasets, all of wafer maps, defect images or question answering, and none of robot handling; a GitHub search for "wafer lerobot" returned no repositories.
Built with what exists
SO-101 leader and follower arms and Hugging Face's LeRobot software with its imitation-learning policies; a 3D-printed thin blade gripper; one open wafer cassette; one or two USB cameras; the Hugging Face Hub. Note: "No download" covers LeRobot v0.7.0.
First step
Assemble and calibrate the arms, use plastic or acrylic discs at first (silicon dummies break into sharp pieces, so wear eye protection), and record 50 teleoperated demonstrations of moving one disc between two slots.
Done looks like
one uncut video of 100 consecutive autonomous transfers, tallied per attempt (success, drop, edge contact, wrong slot, wafer resting across two slots); at least 90 successes and no drops; the dataset and policy are public so anyone can recount from the video.
Open 02 · the brief
Overhead transport scoreboard
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A frozen public benchmark for dispatching and routing overhead hoist vehicles on the SMAT2022 fab layout, run in the open U-FAST simulator, with a scoring script and a leaderboard outsiders can submit to.
Why it is still open
SMAT2022, from Ajou University, ships its layout as Excel files with a demo simulator that "only works in Windows environments"; U-FAST lists baseline scripts but publishes no results table or leaderboard; and an August 2026 routing paper reports results only on its own nine settings, so figures from different papers cannot be compared.
Built with what exists
U-FAST (Python, MIT licence); the SMAT2022 layout and SMT2020, a public model of a fab's production published in 2020; GitHub with automated checks; Gymnasium for learned policies.
First step
Run U-FAST's quick start (one simulated day, 100 vehicles), record the output, and open an issue proposing the frozen scenarios and measures.
Done looks like
at least three frozen scenarios (for example 100, 150 and 200 vehicles, seven simulated days, ten fixed seeds); published mean and 95th-percentile delivery time, deliveries per hour and deadlocks; results for at least three baseline rules; at least one scored submission from outside the team.
Open 03 · the brief
Test AI on GEM manuals
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of interface manuals for imaginary tools, each with a hidden answer key listing the tool's status variables, collection events, alarms and remote commands with their numeric IDs, and a script that scores how well a model extracts them into a fixed JSON format. This scores reading documents, not whether a tool's software behaves correctly.
Why it is still open
Ergotech says these documents "have traditionally required a human to read and interpret them" and publishes no accuracy figures or data; a Hugging Face search for "secs" returns 21 datasets, none about SECS/GEM.
Built with what exists
Python templates rendered to PDF; the open-source secsgem library to keep IDs and message formats valid; open-weight and commercial AI models; a Hugging Face dataset and Space for the leaderboard. Note: "The jobs agents repeat" covers a small model built for extraction and long-document parsing.
First step
Write the answer-key format and one 10-page manual for an imaginary tool, and have one automation engineer check that it reads like a real one.
Done looks like
at least 50 manuals (40 public, 10 held back); an F1 score per field (a standard score that counts both misses and wrong answers); scores for at least three models; a public leaderboard.