Nova
Dispatch

Sony Built the Visor. Siemens Brought the Engineers.

Siemens and Sony used CES 2024 to unveil NX Immersive Designer, which pairs Siemens NX design software with a Sony XR headset and is due later in 2024.

Las Vegas in January is a fever dream with valet parking. I was wedged into the Siemens press scrum, third coffee sweating in my hand, while a man ahead of me in a lanyard explained the industrial metaverse to his phone with the conviction of a tent preacher. Then they rolled out the hardware.

Siemens and Sony had built a thing together. Siemens called it NX Immersive Designer: the company’s NX engineering software, the stuff people use to draw gearboxes and chassis and the guts of machines, fused with a Sony mixed reality headset. The headset had 4K OLED microdisplays and came with controllers for grabbing and turning 3D objects in the air. One was a pointer for precise work. The other was a ring you wear on your finger, which made every engineer in the room look briefly like a stage magician.

They demonstrated it with Red Bull Racing, because of course they did. Nothing says sober product engineering like a Formula 1 team in the booth and a crowd craning to see a car that wasn’t there.

The whole package rides on Siemens Xcelerator, the company’s open digital business platform, and Siemens also used the show to widen its deal with Amazon Web Services, wiring Amazon Bedrock’s generative AI models into its Mendix low-code tools. Availability for the headset system: later in 2024. Price: nobody said.

I walked out into the desert light with my eyes stinging, convinced every gearbox on earth was about to be designed by people waving their hands at nothing, and that the man in the lanyard was still out there somewhere, preaching.

Light Glasses for Heavy Work

Vuzix introduced its 35-gram Z100 smart glasses for enterprise workers at CES 2024, offered first as a $799 developer edition.

Vuzix brought a pair of glasses to CES. They weighed 35 grams. About what ordinary glasses weigh.

There was one small screen, in the right lens. Green. Monochrome. 640 by 480. A MicroLED projector and a waveguide put the picture there. The glasses talked to a phone over Bluetooth, Android or iPhone. You tapped the temple to work them.

They were made for people at work. A part number. A diagram. A direction. A task, confirmed. Data from a scanner, a sensor, a controller on a machine. The lenses were impact resistant and safety-glasses certified. They took prescription inserts. Vuzix said the charge could last up to 48 hours.

The first version was for developers. $799 bought the glasses, the SDK and sample apps.

No headset strap. No visor. Just a pair of glasses on a worker’s face, and a green line of text that tells him which bolt.

Seven Million Parts

Siemens said at Nvidia GTC on March 18, 2024, that it would bring Omniverse Cloud APIs into its Teamcenter X software, with shipbuilder HD Hyundai among the first to use it.

A ship can hold more than seven million parts.

HD Hyundai builds ships. Some of them are meant to run on ammonia. Some on hydrogen. Every part is in a file somewhere. Every file has to agree with every other file.

On March 18, at Nvidia’s GTC conference, Siemens said it would bring Nvidia’s Omniverse Cloud APIs into Siemens Xcelerator. The first product would be for Teamcenter X, its cloud software for managing a product’s life. It would show real-time, physically based digital twins. It was due later in 2024.

Generative AI would help set up the pictures. Materials. Lighting. The scenery around the object. Siemens said work that took days could take hours.

The engineering data would look as it would look in the world. Siemens said that would help sales and marketing teams, and the people who make decisions, and the customers.

HD Hyundai expected fewer errors. Less time. Less cost.

Seven million parts. One picture.

In Which Sailors Learn the Sea Indoors

Danish simulation firm FORCE Technology became an integrator partner for Finnish headset maker Varjo in March 2024, to deliver portable VR and XR training to the maritime industry.

A traditional ship simulator is, roughly speaking, a ship’s bridge built inside a building, surrounded by enormous screens, so that people may practise steering a vessel weighing as much as a small town without leaving a car park in Denmark. It is magnificent, and costly, and you cannot take it anywhere, which is an odd quality in something designed to teach travel.

On March 18, 2024, FORCE Technology, a Danish company that does maritime simulation, engineering and training, signed a strategic frame agreement with Varjo, the Finnish maker of high-end industrial VR and mixed reality headsets. FORCE became Varjo’s integrator partner, which is to say it would take Varjo’s XR-4 Series headsets and wrap its own simulation software around them, then sell the result to maritime customers around the world.

The point was portability. The two companies talked about a compact training setup that could go anywhere, with lower costs and fewer logistics than the traditional sort. FORCE spoke of bringing training out of the classroom and onto ships themselves. The backdrop, they said, was a global shortage of skilled seafarers.

I find it wonderful that the solution to a shortage of people who go to sea is to let them learn it in a headset, standing perfectly still. Somewhere a cadet is about to execute a flawless turn in heavy weather, then take the thing off and discover he is in a portakabin, and the kettle has boiled.

The Yogurt Pot Problem, Solved to 95 Percent

TOMRA showed GAINnext, a deep-learning add-on that separates food-grade from non-food-grade plastics, at the IFAT trade fair in Munich in May 2024.

IFAT, the environmental technology fair held in Munich every two years, is the sort of event where grown adults stand around a machine that sorts rubbish and say ‘ooh.’ I mean that as the highest compliment. From May 13 to 17, 2024, TOMRA set up its booth, B6.339/438 if you were going, under the theme of transforming resource recovery, and gave people plenty to ooh at.

The star was GAINnext, an add-on for its AUTOSORT optical sorters. Its trick is telling apart two pieces of plastic that look, to you and me, entirely alike: one that once held food and one that held, say, engine oil. GAINnext separates food-grade from non-food-grade PET, PP and HDPE at scale, with purity above 95%, using deep learning. TOMRA pointed to coming EU rules on packaging waste and recycled plastic for food contact as the reason this matters.

There was more. AUTOSORT PULSE sorts aluminium alloys using a laser technique called dynamic LIBS, with AI to pull apart objects lying on top of one another, which is a problem I recognise from my sock drawer. TOMRA Insight offered cloud monitoring of sorting lines, and the company mentioned it had taken a 25% stake in PolyPerception, which makes a real-time waste analyser.

TOMRA first applied deep learning to waste sorting in 2019, and to wood in 2022. Five years on, the machines can tell which yogurt pot is fit to become another yogurt pot. My own kitchen recycling still can’t tell a pot from a lid.

Smart Sorting Heads for Tobacco Country

Recycling firm RDS said in May 2024 it would sort single-stream recycling at North Carolina’s Pitt County Recycling Center with a complete AI sortation system from AMP.

Eastern North Carolina, May, the air thick as a wet towel, and I’m parked by a tobacco field with the windows down and a gas-station coffee going warm on the dash, reading a press release from May 6 that says the AI sorters are coming to Pitt County. Not Denver. Not Silicon Valley. Pitt County.

RDS, which owns and runs the Pitt County Recycling Center in Greenville, had won the processing contract in July 2023 and bought the place. Now it said it would fill it with a complete AMP ONE system, AI-driven sortation from AMP, and start processing in June. About 10,000 tons a year of single-stream and commercial recycling. These two companies go back to 2017, and RDS bought its first AI sorter from AMP in 2019 for a plant in Roanoke, Virginia.

Here’s the part that made me sit up. AMP’s founder Matanya Horowitz said that in small and rural markets the costs of hauling and sorting usually kill recycling stone dead, and that this project showed a plant could open profitably in an underserved place anyway. RDS’s president said AI pays off best when you design it into a facility from the start.

A truck rumbled past me loaded with something under a tarp. Maybe hay. Maybe the future. In June it would all start going down the belt.

Foxconn Builds a Factory Twice, Once in Software

NVIDIA used SIGGRAPH 2024 in Denver to unveil generative AI models and NIM microservices for OpenUSD, and said Foxconn was using them to build a digital twin of a factory under development.

There is a peculiar modern habit of building things twice. Foxconn, the electronics manufacturer, was in the summer of 2024 putting up a factory, and at the same time putting up a second copy of it inside a computer, where the forklifts never run out of charge and nobody ever trips over a cable.

The news came out of SIGGRAPH in Denver in the last week of July, where NVIDIA announced a batch of generative AI models and NIM microservices for OpenUSD, the 3D scene format that began life at Pixar and has since wandered into steelworks and assembly halls. Foxconn, NVIDIA said, was using the microservices with its Omniverse platform to build a digital twin of a factory still under development.

The tools themselves have names that sound like a firm of very small solicitors. USD Code answers questions about OpenUSD and writes the Python to go with them. USD Search rummages through big libraries of 3D and image files when you describe what you want in words or show it a picture. USD Validate checks files against OpenUSD versions. Coming later were USD Layout, which assembles scenes from text prompts, USD SmartMaterial, which applies materials to CAD objects, and a service that turns point clouds into meshes.

That last one belongs to fVDB, a new framework that converts photogrammetry, NeRF or lidar data into large environments rendered in real time. NVIDIA claimed four times the scale of earlier frameworks, 3.5x the speed and ten times as many operators, which is the sort of sentence that makes me want to lie down in a darkened room.

Wherever the real building was going up, a concrete floor was still drying. Its copy had been finished for weeks.

Two Thousand Shots a Minute at the Can Line

TOMRA Recycling added a used beverage can application to its GAINnext deep-learning sorter, aimed first at the Americas and rated at up to 2,000 ejections per minute.

I was standing on a steel catwalk above a recycling line on the thirtieth of July, or at least I was in my head, earplugs in, coffee sweating in a paper cup, when the TOMRA release came in and the whole belt below me started to look like a shooting gallery. GAINnext, their deep-learning sorter, had learned a new trick: picking used beverage cans out of an aluminum stream and firing everything else off the line with puffs of air, up to 2,000 ejections a minute. I counted to sixty and tried to imagine two thousand of anything. My heart rate went up out of sympathy.

The machine looks with an RGB camera trained on thousands of images. It knows a drink can by its shape, its size, its dimensions, and it knows an impostor when it sees one: aerosols, food cans, trays, stray metals, plastics. When objects pile on top of each other the software works out where each one actually is and aims accordingly. TOMRA claimed purity above 98 percent and up to 33 times the throughput of a human picker, who manages about 60 picks a minute on a good day with a following wind.

It bolts onto the end of an existing materials recovery line. No new building. No new belt. Just a new pair of eyes that never blink, aimed first at recyclers in the Americas.

I climbed down off the catwalk that wasn’t there. The cans kept coming. Somebody’s soda from last Tuesday, flying straight and true into the right bin.

The Machines Nobody Shipped to Chicago

Chiron Group brought a virtual showroom to IMTS 2024, letting visitors configure its machining centers in 3D on touch screens and project them life-size with Apple devices.

McCormick Place in September is a hangar full of thunder: forklifts beeping, spindles screaming, a thousand salesmen running on hot dogs and adrenaline. I was by the coffee cart with my badge on backwards when I saw the Chiron booth and stopped dead, because something was missing. The machines. Most of them, anyway.

Chiron Group had decided not to ship its whole lineup across the ocean. Instead it brought big touch screens where you could build a machining center in 3D, choosing twin spindles, a five-axis setup, a four-axis pallet changer, tapping and spinning like a kid at a candy counter. And then, the part that made me grab the rail, a salesman held up an Apple device and a life-sized machine appeared on the carpet in front of me, as solid-looking as a parked truck and weighing nothing at all.

Two real ones stood guard so nobody could accuse them of witchcraft. The Micro5, about a square meter of floor and a spindle that runs up to 60,000 rpm, which Chiron pitched as maybe the smallest five-axis machine at the show. And the Mill 2000, a five-axis vertical machining center for big workpieces, topping out at 20,000 rpm.

I walked around the ghost machine twice. Then I put my hand straight through the spindle housing and laughed like a man who had just seen the future and found it cheaper to ship.

The Fee Died. The Factories Never Paid It.

Unity canceled its per-install Runtime Fee for games on September 12, 2024, while saying its non-gaming industry customers were not affected.

For a year game studios had been angry about a fee.

Unity charged it per install. It was called the Runtime Fee. On September 12, 2024, Matt Bromberg, the company’s chief executive, ended it for games.

The subscriptions came back. Unity Personal stayed free, and its revenue ceiling doubled to $200,000. Unity Pro would rise 8 percent, to $2,200 a seat each year. Enterprise would rise 25 percent. The new prices began January 1, 2025.

One line was short. Non-gaming industry customers were not affected.

Those were the others who used the engine. The ones building training simulators, product viewers and digital twins of machines. They had not been part of the fight. They were not part of the peace.

They kept working.

Two Headset Makers Become One

RealWear, which makes wearable computers for industrial workers, acquired Swiss AR headset maker Almer Technologies in a deal strategically and financially backed by TeamViewer.

A technician stands at a machine. Something is wrong with it. The expert is in another country.

This is the job the headsets do. RealWear makes them in the United States. Almer made smaller ones in Bern.

On November 14, 2024, RealWear bought Almer.

TeamViewer backed the deal, with strategy and with money. Its software runs remote support sessions on these headsets. It already owned a minority stake in both companies. It kept that role. Its chief executive, Oliver Steil, took a seat on RealWear’s board.

Almer had rented its headsets by subscription. Compact, simple to use. RealWear brought scale.

Chris Parkinson ran RealWear. Sebastian Beetschen had co-founded Almer and ran it.

Back at the machine, the technician is still waiting. Now there are fewer companies between him and the expert.

A Month of Computing, Squeezed Into an Afternoon

NVIDIA introduced an Omniverse Blueprint for real-time computer-aided engineering digital twins at SC24, with Ansys among the first adopters and claims of up to 1,200x faster simulation.

Engineers who simulate airflow have long had a relationship with their computers rather like the one I have with my plumber: you describe the problem, they go away, and you hear back in a few weeks with a number you don’t fully understand.

At the SC24 supercomputing conference in November 2024, NVIDIA announced something meant to shorten the wait. Its Omniverse Blueprint for real-time CAE digital twins is a reference workflow for makers of engineering software. It stitches together GPU-accelerated solvers from the CUDA-X libraries, physics-based machine learning from NVIDIA Modulus and live 3D visualization in Omniverse, so that when an engineer changes a shape the flow around it updates almost at once.

NVIDIA put the gain at up to 1,200 times faster simulation with real-time visualization, a number I would treat the way I treat the serving suggestion on a cereal box. A more concrete example came from Ansys, one of the first adopters: a 2.5-billion-cell automotive fluid simulation finished in just over six hours on 320 GH200 Grace Hopper Superchips, a job NVIDIA said would take nearly a month on 2,048 conventional x86 CPU cores.

Rescale planned to build the blueprint into its platform. Altair, Cadence, Hexagon, Siemens, SimScale, Trane Technologies and others were evaluating it. Luminary Cloud showed a virtual wind tunnel where you could reshape a car and watch the air rearrange itself.

Nobody, so far as I could tell, demonstrated it on a trommel screen. But the air inside a dust extractor obeys the same equations as the air over a sports car, and it is less fussy about paint.

The Warehouse Before the Warehouse

KION, the forklift and warehouse automation group, showed a digital twin of a robotic warehouse built with Accenture on NVIDIA's Mega Omniverse blueprint at CES on January 6, 2025.

It was a warehouse. It had forklifts that drove themselves. It had cameras in the ceiling and robots on the floor.

None of it was real.

KION showed it at CES in Las Vegas on January 6. KION makes forklifts and warehouse trucks. More than 1.8 million of its trucks were at work at the end of 2023. With Accenture, it had built a copy of a warehouse on NVIDIA's new Mega blueprint, a framework for digital twins of whole industrial sites.

In the copy, the company's own warehouse software handed out jobs. Move these goods from the buffer zone. Put them in storage. The robots that took the jobs were virtual too. They learned the floor. They learned what happens when demand changes and the shelves move.

Operators could try a new layout there. They could try more robots, or fewer people. Nothing stopped on the real floor while they tried.

Accenture built a screen that kept score. Throughput. Time to finish a task. Safety incidents. Errors.

Somewhere a real forklift waited for its orders.

Siemens Hands Engineers a $4,750 Pair of Goggles

Siemens and Sony announced on January 7, 2025 that Sony's SRH-S1 headset and controllers would let engineers edit 3D models directly inside Siemens NX, priced at $4,750 and shipping in February.

Engineers have a long and honourable history of peering at things that are not there. For most of the last century they did it over drawing boards, squinting at elevations and trying to imagine how a gearbox might look if you could walk round the back of it. Then came CAD, which let them squint at a monitor instead. On January 7, Siemens and Sony proposed that they stop squinting altogether and simply put their heads inside the gearbox.

The vehicle for this is Sony's SRH-S1, a headset with 4K OLED microdisplays and a pair of controllers, one shaped like a ring and one like a pointer, which together sound less like engineering kit and more like props from a stage magician. The bundle costs $4,750, a figure I find oddly reassuring, because it is exactly the sort of number that makes a finance department ask for a second meeting, which is how you know a tool is serious.

With it comes NX Immersive Designer, which lets engineers work on 3D models directly in Siemens NX while wearing the headset, and even navigate menus by voice. A sibling, NX Immersive Collaborator, arrived later in January for design reviews, in the same room or from afar, and remote guests join for free. A third app, Immersive Explorer, works with any headset for more casual looking about.

Preorders opened at once and shipping was set for February. Somewhere a drawing board is leaning against a wall in a garage, wondering what it did wrong.

Why Excavator Mechanics Need Waterproof Glasses

Hitachi Construction Machinery, HMS and Hololight announced on March 17, 2025 a rugged, helmet-mounted XR smart glasses design that lets field technicians get live help from a service desk.

If you have ever tried to read an instruction manual while holding a spanner in one hand and a torch in your teeth, you will understand the problem Hitachi Construction Machinery set out to solve. Its field technicians service excavators and mining machines, often in mud, rain and the sort of dust that gets into sandwiches, and they had been juggling status reports, service manuals and parts catalogues on phones and tablets with hands that were, quite reasonably, already busy.

So on March 17, in Tokyo, the company announced a new design of XR smart glasses, developed with HMS and the German streaming specialist Hololight. The glasses clip onto a safety helmet and carry an IP65 rating, which in plain English means that dust cannot get in and a jet of water will not trouble them. I find this deeply admirable. I own an umbrella that cannot make the same claim.

The technician works with physical buttons and finger gestures while several virtual screens float in the space around him, streamed in by Hololight's technology. The camera sends what he sees back to a service desk, where someone with a clean desk and a hot coffee can offer advice, documents and data. Hitachi says the aim is faster jobs at a time when trained service people are in short supply.

The glasses were due to make their public debut at the bauma exhibition in Munich, April 7 to 13, a trade fair devoted to very large machines and, presumably, a great many very small screws.

Jet Engines in the Break Room

TeamViewer announced Frontline Spatial, an AR tool for industrial training built on interactive digital twins, ahead of Hannover Messe in late March 2025, citing GE Aerospace as a user.

The press release landed on March 26 and I was in the back seat of a cab crawling down the BQE in sleet, laptop balanced on my knees, the driver arguing with a radio host about the Mets, and I read it twice and then once more out loud, which the driver did not enjoy.

TeamViewer, the remote access company everybody's uncle has installed at least once, was walking into Hannover Messe with something called Frontline Spatial. Interactive digital twins of industrial equipment, delivered to technicians through augmented reality, for training and onboarding. Not slides. Not a video. The machine itself, floating in front of you, ready to be taken apart.

And the customer they put on stage was GE Aerospace, a joint customer with Siemens, which uses digital twins of its commercial jet engines to train aircraft technicians remotely across hundreds of locations. Jet engines. Hundreds of locations. I pictured a mechanic in a hangar somewhere pulling a turbine apart in midair with his bare hands while the coffee went cold beside him.

TeamViewer pitched it squarely at the skills shortage: learn remotely or on site, travel less, spend less. On March 31 its CEO, Oliver Steil, was booked on a panel at the Siemens booth with Tony Hemmelgarn, president and CEO of Siemens Digital Industries. TeamViewer's own booth would show a Formula One engine, because apparently a jet engine was not enough.

The cab hit a pothole. The laptop jumped. On the screen an exploded engine hung in the air, every bolt numbered, and for a second I could have sworn it moved.

Twelve Seconds Off Every Robot Arm

Taiwanese manufacturers including Foxconn, Wistron and Pegatron reported gains from NVIDIA Omniverse factory digital twins around Computex, according to an NVIDIA post published May 18, 2025.

It was past midnight on May 18 and I was at my kitchen table in Astoria with the overhead light buzzing and a pot of coffee that had gone from beverage to tar, scrolling through NVIDIA's Computex roundup, and the numbers kept coming at me like rivets from a gun until I had to stand up and walk a lap around the table to keep my heart where it belonged.

Foxconn was designing robot work cells, assembly lines and entire factory layouts inside its own digital twin platform, and running thermal simulations of its server rooms 150 times faster. Wistron simulated robotic arms before bolting a single one to the floor and shaved 12 seconds off each arm's assembly time on the line. Twelve seconds. Multiply that by a factory that never sleeps and you can hear the money falling out of the ceiling.

Pegatron said its visual AI agents had cut labor costs by 7 percent and assembly line defects by 67 percent. Techman Robot trained cobots in simulation at Volkswagen's Transparent Factory and cut programming time by 70 percent. TSMC, the chipmaker the whole planet leans on, was turning 2D designs into interactive 3D fab layouts and generating multilevel piping in seconds.

These are the companies' own numbers, reported through NVIDIA, and I wrote that down twice on the back of an electric bill. Then I looked at the bill. It was the only thing in the room nobody had simulated.

The Glasses Learn the Manual

Vuzix and German startup Ramblr paired video-understanding AI with Vuzix M400 smart glasses to give frontline workers spoken, step-by-step guidance built from a company's own manuals and videos.

The press release went out on May 29. Two companies. One pair of glasses.

Vuzix made the glasses. The model was the M400. It had a small screen and a camera.

Ramblr made the software. It was a German company and it watched video. It tracked objects, actions, and where things were in relation to each other.

You fed it what the company already had. Manuals. Training videos. The way a job was supposed to go.

Then the worker put on the glasses and started the job. The glasses saw the job. The worker heard what came next, in plain language, and saw it on the small screen. Nobody had to stop and look anything up. When the job was done, the glasses could take the photos for the record.

The release gave one number. Seventy percent accuracy on the dynamic, context-based tasks.

It did not say how they measured it.

Seventy percent.

The other thirty were still the worker's.

My Hands Belong to a Stranger in Göppingen

TeamViewer said on July 30, 2025 that analyst firm PAC had named it the only best-in-class vendor for connected worker platforms in its European RADAR report, for the second year running.

I was sitting on an overturned milk crate behind a print shop in Long Island City, eating a gas-station sandwich in the heat, when the press release came in from Göppingen, a German town I could not have found on a map with a gun to my head. TeamViewer had been named the only best-in-class vendor in PAC’s European RADAR for connected worker platforms. Second year in a row. The analyst said the company was growing faster than the market and grabbing share.

Connected worker. Say it slowly and it sounds like a horror film. Here is what it means. A technician on a factory floor puts on smart glasses, or picks up a phone, and a remote expert somewhere else sees what he sees and draws arrows on the machine in front of him. Step-by-step workflows float in his eyeline. TeamViewer Frontline plugs into SAP and Siemens systems, and a module called Frontline Upskill trains technicians on 3D models of engines and aircraft parts. There is AI now too: live multilingual subtitles, and workflows generated straight out of a PDF manual.

The numbers came at me like hail on a tin roof. Volvo Group, picking quality up 25% at a Sydney distribution center. Vandemoortele, training time for temporary workers down 25% across six warehouse sites. Around 660,000 customers. Around €671 million in revenue in 2024.

I finished the sandwich and looked at my own two hands for a long time, wondering whose instructions they would be following next year.

Panorama, Then Mesh

Tencent released HunyuanWorld 1.0 on July 27, 2025, an open model that turns text or images into explorable 3D environments and is licensed for commercial use.

Tencent put it out on the last Sunday of July.

You type a sentence. Or you give it a picture. It makes a world.

First it draws a 360-degree panorama. Then it takes the panorama apart. Then it builds it back as a layered 3D mesh you can walk through in real time.

It runs on the Flux image model. It can be adapted to others, Stable Diffusion among them.

It is free. The code is on GitHub and Hugging Face. The license allows commercial use. There is a web version for those without the machines to run it.

Tencent did not say who it was for. It said games. It said virtual experiences.

Somewhere a designer has a plant to lay out. A hall. A floor with room for a shredder and a belt and a baler. She has drawn that floor by hand before. She has drawn it many times.

Practice on the Real Machine

Augmentir added an augmented reality module to its connected worker platform on September 23, 2025, aimed at onboarding, training and guidance for complex or infrequent tasks in manufacturing.

The announcement came out of Horsham, Pennsylvania, on September 23. Augmentir makes software for people who work on factory floors. Now it had added augmented reality.

The idea is simple. A new hire stands in front of the real equipment. The instructions appear on top of it. She practices on the machine she will run, not on a picture of it. When a job comes along that nobody has done in months, the steps are there again, laid over the parts.

The AR module plugs into the rest of the platform and its AI agents. The AI shapes what the guidance shows. The company had built its first mobile AR prototypes in 2020. It waited five years to sell it.

Augmentir said AR earns its keep in training, in onboarding, in the first time a task is done. Not every day. The first time.

The release listed results its customers had reported from the platform as a whole, not from AR alone. Onboarding time down 82%. Downtime down 27%. Safety incidents down 21%. Colgate-Palmolive was on the customer list. So was Mondelez. So was AkzoNobel.

The first time.

A Recycling Robot Buys the Building It Lives In

AMP, the AI waste sorting company, completed its purchase of RDS of Virginia’s Portsmouth operations in late September 2025, including a single-stream recycling facility it plans to expand.

There is a particular kind of tenant who loves a flat so much that one day they simply buy it from the landlord. I have never been that tenant. I once lived in a basement in Leeds where the hot water arrived with the punctuality of a comet. But AMP, the AI sorting company, appears to have done exactly this in Portsmouth, Virginia.

Since late 2023, AMP’s system, called AMP ONE, had been running at an RDS of Virginia facility in Portsmouth. It takes ordinary bagged household trash, the kind nobody has bothered to sort, and pulls the recyclables and the organics out of it, at up to 150 tons a day. In late September AMP announced it had completed its purchase of RDS’s Portsmouth operations, including a single-stream recycling facility that it intends to expand.

One hundred and fifty tons is worth dwelling on. That is roughly the weight of a blue whale, sorted, every day, by a machine that does not need a tea break. I cannot sort my own sock drawer at that rate, and my sock drawer is mostly one sock.

AMP’s chief executive, Tim Stuart, said the company would keep looking at acquisitions that let its systems handle more tonnage, and described Portsmouth as a model other municipalities could copy. The landlord, presumably, has moved on to other basements.

Rimini, Where the Rubbish Gets the Red Carpet

More than 100,000 visitors attended Ecomondo 2025 in Rimini from November 4 to 7, where Spanish firm PICVISA showed AI and hyperspectral sorting machines for textiles, glass and e-waste.

Rimini is best known as a beach resort, the sort of place where Italians go in August to lie in neat rows under umbrellas arranged with military precision. Every November, however, it hosts Ecomondo, a trade fair for the green and circular economy, and the umbrellas are replaced by optical sorters, balers and people in lanyards discussing polymer recovery over very small coffees.

The 2025 edition, the 28th, ran from November 4 to 7 at the Rimini Expo Centre. More than 100,000 visitors came, there were over 200 events in the conference rooms, and 40 start-ups took part. That is a remarkable number of people to gather in one place to look at machines that look at rubbish.

Among the exhibitors was PICVISA, a Spanish company that builds sorting machines using artificial vision, robotics and AI. Its textile sorter, ECOSORT TEXTILE, combines hyperspectral imaging with AI and now has robotic feeding. The company pointed to Lipor in Portugal as the first automatic textile sorting plant, to a turnkey glass project in Brazil whose AI spots dark or labelled glass, and to new lines in the United States for things such as shoe soles and shredded electronic waste. It also showed an optical line, built with Eldan Recycling, for cables, e-waste and metals.

I have always liked the idea that somewhere a machine is looking very carefully at an old shoe sole and making a considered decision about its future. It is more thought than I give most of my shoes.

A trade show remains the one place a buyer can stand next to a machine the size of a minibus. That is a stubborn fact for anyone selling one, and an argument for bringing a 3D version that fits in a suitcase.

Microsoft Walks Out of the Room and Leaves the Lights On

Microsoft stopped selling and renewing Dynamics 365 Guides and Remote Assist on November 1, 2025, ahead of retiring both mixed reality apps on December 31, 2026.

November 1. A Saturday. I was in the passenger seat of a borrowed pickup on the Gowanus Expressway, rain smearing the windshield, the wipers losing the fight, a paper cup of coffee wedged between my knees, when a field engineer I know called to tell me the door had closed. As of that day, nobody could buy or renew Dynamics 365 Guides or Dynamics 365 Remote Assist. Microsoft’s mixed reality workhorses. The step-by-step holographic instructions. The calls where an expert three time zones away could see through a technician’s headset and point at the right valve.

Remote Assist had been around since September 2018. Now both apps have a date with the end: December 31, 2026, when the security updates stop. Existing subscriptions keep working until then. After that you are running unpatched software on a factory floor, which is like leaving the loading dock door propped open with a brick and a sign that says welcome. Microsoft’s advice to customers is to look at partner products in its marketplace. In October 2024 it had already said there were no plans for a new HoloLens.

The engineer on the phone had trained half a plant on those headsets. He did not sound angry. He sounded like a man reading the small print on a lease. The rain got harder. A truck passed us throwing up a wall of grey water, and for a second the whole highway disappeared, and then it came back, exactly the same, except it didn’t feel like it.

PepsiCo Rehearses Its Factories Before Touching a Bolt

Siemens launched Digital Twin Composer at CES 2026, and PepsiCo reported a 20% throughput gain from using it to model upgrades at US plants and warehouses.

There is a long tradition in my family of assembling flat-pack furniture without reading the instructions, then reading the instructions, then going to buy a different wardrobe. PepsiCo, it turns out, has taken the opposite approach to its factories.

At CES in Las Vegas in early January, Siemens unveiled Digital Twin Composer, its headline launch for the show. The software stitches together Siemens' own digital twin tools, simulation built on NVIDIA Omniverse libraries and live engineering data, so a company can build a 3D model of a product, a process or a whole plant and then scrub forward through time to see what a change will do. It is due on the Siemens Xcelerator Marketplace in mid-2026.

PepsiCo has been using it to turn selected US manufacturing and warehouse sites into 3D twins, with machines, conveyors and the paths operators walk all simulated before anyone picks up a spanner. The company's own figures are the sort you would normally expect from a man selling miracle tonic at a county fair: up to 90% of potential issues caught before physical changes, a 20% throughput increase on the first deployment, and capital spending down 10 to 15%. Plans are to scale it globally.

In the same announcement, Siemens said it was bringing its industrial AI to Meta's Ray-Ban AI glasses, to whisper guidance and safety notes into the ears of shop floor workers.

Somewhere, I imagine, a wardrobe in a Siemens simulation is being assembled perfectly on the first try.

Yellow Iron Learns to Think in the Desert

Caterpillar used CES 2026 to describe factory digital twins built on NVIDIA Omniverse and to preview five autonomous construction machines running NVIDIA AI.

Las Vegas, January 8, and I am standing in a convention hall the size of a small weather system with a paper cup of coffee that tastes of burnt cardboard and ambition, staring up at a Caterpillar machine so large its shadow has its own postcode. Somebody near me says it can drive itself. Somebody else laughs nervously. I do not laugh. I write it down.

Here is what Caterpillar actually said, and I checked it twice. It is building digital twins of its own factories on NVIDIA Omniverse and OpenUSD, virtual copies where engineers design and tune production processes before a single thing moves on the real floor. It previewed five autonomous construction machines, drawing on three decades of driverless mining work. Future machines will carry NVIDIA's Jetson Thor for real-time AI, for autonomy and for nudging human operators with safety alerts.

The numbers came at me like hail. A Cat AI Assistant sitting on a data platform called Helios holding more than 16 petabytes. Off-board features from Q1 2026, in-cab ones in final validation. Thirty billion dollars in R&D over two decades. Digital and technology spending to rise 2.5 times by 2030. A $25 million prize over five years for workforce education.

I called my editor from the parking lot. The sun was setting purple over the Strip, and somewhere in a server rack a whole Caterpillar factory was being rebuilt, line by line, in a world nobody can stand in. She asked if I was all right. A bulldozer the color of a school bus rolled past the window of my rental car with nobody in the seat.

Strap In, Kid, the Dozer Is in Your Head

John Deere introduced a portable Extended Reality Training System for its 650 P-Tier dozer and 210 P-Tier excavator at CONEXPO-CON/AGG 2026 in Las Vegas.

Silver Lot at CONEXPO, Las Vegas, March, wind whipping grit off the desert into my teeth, and a Deere rep hands me a headset and points at a folding chair like it is the electric throne, and I sit, and the parking lot vanishes, and I am in the cab of a 650 P-Tier dozer with a blade the width of a garage door and dirt rolling toward me like a brown tide, and my hands are sweating on controllers that are not levers but my brain insists they are levers, and the rep is saying something about augmented and virtual reality in the same system, operator lessons and maintenance lessons and even sales training, and I am pushing dirt, actual imaginary dirt, and the 210 P-Tier excavator is waiting for me after this one, and the hardware is just a Meta Quest 3 or a Pico 4 Ultra Enterprise, the kind of thing you could carry into a dealership or a classroom or a jobsite trailer under one arm, and later I will read in Equipment World that the single-headset bundle runs about $4,000 with renewals near $2,600 while the old seated simulators cost something like $85,000 to $100,000, and I will sit down hard on a curb doing that arithmetic, but right now the rep taps my shoulder, my time is up, there is a line of men in hi-vis behind me, and I take the headset off and the sun is enormous and the real dozer on the stand looks at me like it knows.

First the Model, Then the Concrete

Nvidia made its Omniverse DSX Blueprint generally available at GTC, letting engineers simulate an AI factory's power, cooling and operations before construction.

There is a field in West Virginia. Grass. Fence posts. A road. Wind.

Nscale and Caterpillar plan a multi-gigawatt AI factory there.

On March 16, at GTC in San Jose, Nvidia made its Omniverse DSX Blueprint generally available. A digital twin of the whole plant. Power. Cooling. Networking. Operations. Layouts, heat, policies, all in one model.

Engineers test it first. Then they build.

Partners signed on. 17 of them. Thirteen from industry, four from energy. Siemens. Schneider Electric. Dassault Systèmes. Vertiv. Trane. GE Vernova. Hitachi.

Phaidra said its cooling work could add about 10% more compute.

More than 200 gigawatts of US projects wait in line for power.

The field waits too.

A Sorter Learns the Copper Meatball

TOMRA introduced new deep learning sorting applications and an AI waste-analysis platform at IFAT 2026, and raised its stake in PolyPerception to 51%.

Scrap comes down the line. Steel. Copper wound around steel. Cans. Trays.

In May, at IFAT, TOMRA showed new deep learning applications for its GAINnext sorters.

One tells food-grade PET trays apart by shape and use. TOMRA put purity above 95%.

One finds the copper meatballs, copper and steel fused together, and pulls them out. Rebar-grade scrap becomes premium furnace feed.

One sorts used beverage cans. Up to 33 times the throughput of hands. Purity 98% or higher.

TOMRA also took its stake in PolyPerception to 51%. Their Waste Analyzer now answers questions in plain language. It writes quality reports. It sets alerts. Show it a battery and it finds other things that look like batteries, without new training.

The cans keep coming. The machine keeps looking.

Spectacles for Everyone Except, So Far, the Night Shift

Google and Samsung showed their first Android XR smart glasses at Google I/O on May 19, 2026, and Xreal said its Project Aura glasses would go on sale before year end.

The history of computers you wear on your face is long, earnest and mostly embarrassing, rather like the history of the moustache. So I tuned in to Google I/O on May 19 with the mild dread of a man watching a relative take the microphone at a wedding.

It went better than feared. Google and Samsung showed their first Android XR glasses, with audio models due in fall 2026 in select markets and frames from Gentle Monster and Warby Parker. These ones talk to you through Gemini, offering directions, messages and live translation of signs and menus. Glasses with a display in your view come later, with no date, name, price or battery life announced, which leaves a good deal to the imagination.

Xreal, meanwhile, brought Project Aura: wired glasses with optical see-through OLED displays and a 70-degree field of view, promised for sale before the end of 2026. Google and Xreal also opened a developer programme lending early hardware to selected builders.

None of this was pitched at the factory floor. But I could not help remembering that back in January, Siemens said it would bring its industrial AI to Meta's Ray-Ban glasses, audio first, murmuring troubleshooting steps and safety notes into the ears of shop floor workers. The consumer world spent May arriving, with great ceremony, at roughly the place a maker of motors and switchgear had pencilled in over the winter.

I find that pleasing, in the way it is pleasing when the tortoise turns out to own a car.

I Strapped On the Warehouse and Waited for Glory

A controlled study of 100 trainees, published online July 15, 2026, found VR order-picking training on Meta Quest 3 made people more confident but did not significantly beat a written guide on speed or errors.

I read the paper at a folding table in a borrowed office in Queens, the air conditioner wheezing like an old forklift, a cardboard cup of bodega coffee sweating rings onto page nine, and I was ready, I mean truly braced, for the moment virtual reality crushed the printed instruction sheet into pulp. The study came out July 15 in the Flexible Services and Manufacturing Journal, by Adamos Daios and Ioannis Kostavelis of the International Hellenic University in Katerini, Greece, and it was built like a proper fight: 100 undergraduates, split down the middle, fifty learning manual order picking from a written step-by-step guide and fifty getting about eight minutes inside a virtual warehouse built in Unity and run on Meta Quest 3 headsets. Then everyone went and picked orders for real. I turned the page. The VR group averaged 321.56 seconds. The paper group averaged 334.90. Not significant. The VR group averaged 2.04 errors to the paper group’s 2.88. Close, tantalizing, still not significant. I put the coffee down. What VR did win, and win clearly, was the head game: its trainees felt faster, more effective and more confident, scoring 4.46 on confidence against 4.06, and ten days later they still expected to improve more, even though their recall was no better than the paper crowd’s. Somewhere a training manager was already ordering forty headsets. Somewhere else, a laminated sheet of instructions sat on a shelf, curling a little at one corner, undefeated.

The Mine Changes Every Day

Caterpillar has acquired Skycatch, a company that captures and processes high-precision spatial data at mining and industrial sites, The XR Beat reported July 9, 2026.

A mine is never the same place twice. Trucks take the ground away. Benches move. Stockpiles rise and fall.

The map is always behind.

Caterpillar bought Skycatch, The XR Beat reported on July 9. Skycatch captures high-precision spatial data across mining and industrial sites and processes large sets of it at speed. The point is to shorten the time between a change in the ground and a change in the plan.

Caterpillar already owns RPMGlobal and its MineStar software. Skycatch’s near-real-time, high-resolution spatial data will go into both. No price was given. No terms. No figures for speed or accuracy.

Caterpillar called it near-real-time. That is the whole idea.

The ground keeps moving. Now the map tries to keep up.

A Robot With a Clipboard, Somewhere in Texas

Caterpillar and robotics startup FieldAI announced a collaboration on September 2, 2026, to develop autonomous inspections and jobsite and facility digital twins using NVIDIA Omniverse.

Caterpillar has been making yellow machines for, by its own count, more than a century, which is long enough that the company predates the traffic light, the zipper as we know it, and the general expectation that a bulldozer might one day have opinions. On September 2, from Irving, Texas, it took another step toward that last one.

The company announced a collaboration with FieldAI, a startup that builds AI foundation models for robots. FieldAI’s models are robot-agnostic, which is to say they are meant to work on whatever robot you happen to have lying around, a quality I also look for in batteries. NVIDIA supplies accelerated computing and its Omniverse technologies.

The early to-do list is practical rather than cinematic: autonomous inspections that keep people out of riskier spots, digital twins of jobsites and facilities built from operational data, sharper situational awareness to catch problems sooner, and simulation to tune how work flows. Caterpillar said the same tools will be turned on its own factories, to better understand production flow and safety. FieldAI’s chief executive, Ali Agha, says the company is already deployed at hundreds of sites around the world.

No prices, no timetable and no financial terms were disclosed, which leaves me free to imagine the eventual result: a mining site in which a robot strolls past an excavator, takes careful notes, and files them into a perfect virtual copy of the very ground it is standing on, while the excavator, I assume, pretends not to notice.

Seventy Percent Have No Screen

Audio glasses made up 70.3% of XR headset and glasses shipments in the second quarter of 2026, IDC reported September 16, as total shipments grew 35.3% year over year.

IDC published its second-quarter count on September 16.

Shipments of XR headsets and glasses rose 35.3 percent from a year earlier. The quarter before, growth had been 130.2 percent.

Audio glasses took 70.3 percent. Speakers. A microphone. A camera. An assistant. No display. A year earlier they had 61.9 percent.

Display glasses took 14.3 percent, up from 11. Headsets took 15.4.

Meta held 68.7 percent of the market. IDC said its lead came largely from distribution.

Behind Meta came RayNeo, Alibaba, XREAL and Xiaomi. Small numbers. Each under four percent.

IDC expects unit growth of 26.3 percent for the full year.

RayNeo grew fastest of the rest. Up 83.3 percent. It held 3.6 percent of the market.