Dear Aventine Readers, 

Headlines about driverless taxis, warehouse bots and the occasional barista bot can create the impression that robots are everywhere, and that pretty much anything that can be automated is. This turns out to be pretty far from the truth. Incorporating robots into a workplace can be difficult and costly, often requiring changes to physical spaces, production plans and staffing. But, as we explore in this issue, a number of smallish, nimble companies are trying to change that, taking advantage of a confluence of factors to build more affordable, low-hassle robotic systems for customers who might never have considered automating before and might be in the market for just a single robot. Read on to learn about how these companies are transforming the process of automation. 

Also in this issue: 

Thanks for reading as always, 

Danielle Mattoon 
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

At the end of a plumbing hardware production line in Glasgow, Scotland, sits an employee with a single task.

The person picks components out of a box containing a collection of parts made on the factory floor, turns around, and places the parts down precisely so that they can be measured for quality assessment. Components that pass the test continue their journey to be packed; those that fail are discarded.

The company, McAlpine & Co Ltd, decided that this person’s time could be better spent doing something else at the factory, so it decided to explore the idea of automating the process with a company called Vention. Within a few weeks, Vention had designed a proof-of-concept system using a robot with an AI-based vision system and finger-like grippers to pick and organize components. The system is now in testing, and will soon be tried out on the McAlpine factory floor.

Vention is just one company among plenty of others across Europe and North America taking advantage of a confluence of factors to build more affordable, low-hassle robotic systems for customers who might never have considered automating before. At a time when cutting-edge, AI-enhanced robots are still largely confined to research labs, existing robot technologies can be put to use more cheaply and efficiently than ever, thanks to falling hardware prices, new business models and innovation in everything from open-source AI models to the ability to create digital renderings of factory sites. 

"How do we drive massive accessibility in robotics and automation so that you get adoption at scale?" asked Saman Farid, CEO and cofounder of Formic, a startup that rents out robots designed to fulfil specific tasks in industrial settings. "Part of that is a business model question, part of that's a technology question, part of that is an operational question." If companies like Formic find success, they could unlock a new stream of automation in areas of manufacturing that have traditionally been reluctant to bring on robots. 

The automation lag

A paradox of the current age of robotics is that even though we live in an era ripe for automation, with manual labor shortages and high turnover in warehouse work, industry hasn’t automated as fully as one might think, given the amount of attention given to autonomous cars and Amazon warehouse bots. According to the International Robotics Federation, it’s taken almost seven years for global robot density in factories to double, from 74 robots per 10,000 employees in 2016 to 162 in 2023. While there was a surge in robot investment during the Covid-19 pandemic when businesses panicked about maintaining production levels with reduced workforces, purchasing levels have since fallen sharply. 

Adoption also happens to be dominated by just a few industries. Of the 4.3 million industrial robots that the IFR estimated to be in service around the world in 2023, 30 percent are used in the automotive sector, a further 28 percent in electrical and electronics, and the remaining 42 percent shared across all other industries. In all these industries, adoption is “really concentrated,” in the sense that a small number of companies have invested heavily in the technology while others haven’t at all, said Christopher Müller, director of the IFR’s statistical department. 

Small- and medium-size companies in particular have been reluctant to invest in automation, said Jeff Burnstein, president of the Association for Advancing Automation (A3), a trade group representing companies involved in robotics. Several factors account for that, he explained: They’re reluctant to take on, or simply can’t afford, the large up-front expense of robotization; they fear the disruption it may bring; and they don’t have staff members with the skills required to make it happen. “A lot of those small companies don't have those internal resources,” he said.

A new kind of robotics company

Historically, the robotics industry wasn’t really set up to help small businesses, or larger ones looking to experiment with only a small number of robots. “If you look at it from the standpoint of the big industrial robot players … going to find every little guy who is going to buy one [robot] — that was not part of their plan,” said Burnstein. Even robotics integrators, companies that can customize robots to fit the needs of a company, “aren't really motivated, necessarily, to go find onesies or twosies,” he added. But that has begun to change, with an increasing number of businesses founded on the notion that it should be possible to change the way companies start using robots in order to quicken adoption. “You don't want to buy a robot, you want to solve a problem,” said Müller. “I think this is something that the industry has been learning over the past few years, and many companies are approaching this now.” 

Vention, for instance, founded in 2016, specializes in working with clients to design robot systems quickly on specially created software that works in a regular internet browser. This allows customers to specify and assess how a robot and its surrounding infrastructure, such as conveyors and protective cages, can be integrated into a workspace before committing to a purchase. Vention then delivers all the required hardware and software and installs the system; it can also provide ongoing maintenance and even financing, acting as a one-stop-shop for automation. Meanwhile, Formic, founded in 2020 in Chicago, and Robust.AI, founded in 2019 in San Carlos, California, are two of many companies that now provide robots on a rental basis, what’s referred to as robotics-as-a-service, or RaaS. Clients take out a contract, paying for the robot each month like they would for a member of their workforce. 

Part of what has enabled startups to enter this market is decreasing costs. “The cost of robot hardware is dropping precipitously,” said Farid. An average industrial robot cost more than $68,000 in 2005; that is projected to fall to under $11,000 this year. He attributes that reduction to the fact that, first the smartphone industry and then the drone and autonomous vehicles industries, drove down the cost of chips, motors, cameras, sensors and many other kinds of electronic components used in robotics. “For every type of component that would go into a robot, there used to be maybe one supplier, and now there's like 50,” he added. Reduced costs have made the economics of selling or renting even single robots to clients financially viable. Formic, for instance, will happily rent a single robot designed to load palettes to a company for around $5,000 per month. “Think of us like a staffing agency, except we staff with robots,” said Farid. “We'll show up with the robot. It'll do the job, whether it's for one month or 12 months or five years.”

Some previously time-consuming elements of deploying robots, such as design and programming, are also becoming more efficient, making it possible to ship systems to customers faster than before. Etienne Lacroix, CEO of Vention, said that enabling customers to design systems in a simple browser window, rather than with complex computer-aided design software, makes it far easier for clients to see what a system might look like and cost in real time. The McAlpine production-line robot mentioned above, meanwhile, makes use of off-the-shelf robotics chips from Nvidia and open source artificial intelligence software from Meta. “While we could develop our own, we're leveraging what other people are already putting out there,” said Sarah Webster, chief marketing officer of Vention. That allows the company to build robots that perform bespoke tasks more quickly, she said. Lacroix expects systems that make use of AI to become increasingly common at Vention over the next 12 to 18 months. 

Streamlining automation

Limiting the scope of what robots do has also helped these companies streamline their processes to make their business models viable. Robust.AI, for example, founded by a longtime roboticist, MIT professor emeritus and serial entrepreneur Rodney Brooks, is building robots to help warehouse workers locate and transport items and typically offers only one or two products for specific tasks. One such offering is an autonomous trolley that can roll along aisles, indicate where items are to a warehouse employee, and then take itself to a packing desk when it's fully loaded. Vention and Formic, which supply robots for a wide range of applications, also tend to limit themselves to specific tasks that are standard in many lines of business: packing boxes, picking items from conveyors, loading palettes and so on. Some of these are big, industrial robot arms reminiscent of those you might find in automobile factories; others are smaller cobots, which are limited in power and speed so they can safely work alongside humans. Focusing on relatively common tasks means less customization — and therefore less cost — associated with each deployment.

Farid explained that Formic has also automated other processes that it initially found to be expensive or time-consuming. For example, because a site visit for scoping a new installation could take days or weeks of mapping the space and taking measurements, his team developed computer vision tools and other technologies to do the job much faster. When it comes to designing a robot setup, which might require hundreds of decisions, his team now uses software that automatically makes many of those decisions for them. The company has also developed remote monitoring and diagnosis tools to track the performance of robots and often to fix them remotely when they malfunction, reducing the need for maintenance teams to rush out to client sites. Vention, meanwhile, has software that allows customers to build a digital twin of each robot system, so that the clients can assess how it will work on their factory or warehouse floor and iron out any wrinkles before committing to a capital investment.

In contrast to the cutting-edge advancement of robotics taking place in well-financed laboratories around the world, these sorts of startups are focused on using robotics approaches that largely already exist while reducing costs and finding efficiencies. “It isn’t sexy,” said Brooks with a laugh. But it does seem to work: Both Formic and Vention say that they can now move from an initial conversation with a client through design, specification and programming robots, to deploying them in a company’s premises within days or weeks, a process that may historically have taken months, if it ever happened at all. 

“It's much easier now, much cheaper, much faster, to get something that's proven [to provide return on investment],” said Burnstein. 

Waiting for liftoff

To date, none of these startups has brought about explosive growth in automation — the macro trends recorded by IFR reveal as much. But some are getting notable traction. Vention, which serves companies of every size, says that it has deployed more than 4,000 automation systems, and that it now ranks as a top-three seller of cobots in the United States. Formic hasn’t publicly shared numbers of its installations, but says it expects to deploy four times as many robots in 2025 as in 2024, and claims that it has a renewal rate of 98 percent — meaning that almost all of its customers reach the end of the contract and decide to keep using the robot.

Lacroix admitted that automation still needs to become more affordable and easier to deploy if it's going to achieve higher levels of adoption. The robot installations that Vention sells to its clients have an average payback time — how long a capital expense takes to pay for itself — of around 1.3 years, according to Lacroix, but he compared this to payback times of under a year for, say, software purchases. "If a company has only X amount of capital they can invest in a year, they might decide, 'Maybe I'll buy that software, versus hard automation,'" he said. "Automation needs to get below one year payback. We're very close."

Indeed, there’s a sense across the sector that things are on the cusp of changing. “I think we are at a turning point in terms of the adoption of robotics,” said Farid. “It’s been one of those things [that] has been promised for the last 30 years, that robots are going to be everywhere. [But] it does feel like now, finally, we are really at a turning point.”

A heap of electronic and computer hardware waste for recycling Mindful Media/IStock

Recycling electronics could help solve a scarcity of rare earth minerals. China controls about 90 percent of the world’s capacity for processing 17 of the planet’s rare earth minerals — the ones underpinning the electric economy in clean energy technologies like electric motors and wind turbines. As the world braces for a trade war, this asymmetry is prompting some creative solutions. The Financial Times reports on how several startups are developing ways to extract rare earth metals, such as Cerium and Neodymium, from old electronic devices by breaking down the magnets used inside items as diverse as hard disk drives and MRI scanners. One company, HyProMag, based in Birmingham, U.K., puts electronic waste into a large, sealed metal tank and then pumps it full of hydrogen. Tiny atoms of the gas infiltrate cracks in the magnetic materials, causing them to break down. That creates a fine dust composed mainly of rare earths that can then be collected and used to make new magnets. Another company, Cyclic Materials, based in Toronto, Canada, breaks up similar waste into small pieces and then uses chemical processes that can separate materials into their constituent elements, which can then be repurposed. Other companies are developing alternative techniques. Currently, the amount being recycled is low: Cyclic produced just 100 tonnes of rare earths in 2024, but plans to increase that to 600 tonnes this year. HyProMag produces even less. Global demand for rare earths is expected to reach 304,678 metric tonnes by 2025, so these amounts are a drop in the bucket. That said, the approach suggests a new path toward storing up these crucial elements. 

We’re beginning to understand the hard limits of LLMs. Large language models are impressive, right up until they fail spectacularly. Understanding how and why they sometimes fail is important, partly so that we can understand what they can be reliably used for and partly to know where to focus efforts to overcome their flaws. Quanta reports that there’s an emerging field of research trying to determine the precise limits of when a problem becomes too complex for an LLM to solve. One important example is around so-called compositional problems — puzzles that demand facts be pieced together like, say, “Alice lives in a blue house” and “The favorite drink of someone in any blue house is coffee” to answer a question such as “What is Alice’s favorite drink?” Early LLMs were spectacularly bad at this sort of puzzle, and one particularly interesting recent result from UC Berkeley has proven that for transformer-based LLMs — which is to say, most of them — it is always possible to extend a compositional problem so that it is too complex for the AI model to solve. To some extent it may be possible to augment transformer-based LLMs to help them solve such problems, and that phenomenon is perhaps best exemplified by the growing crop of reasoning models developed by AI companies. (DeepSeek’s R1 reasoning model, for instance, solved the well-known Zebra Puzzle for Aventine in four minutes.) Such bolt-ons will likely take LLMs only to a certain level of complexity, according to the experts who spoke with Quanta, and not deliver the intrinsic skills necessary to be error-proof. The larger point is that it’s increasingly possible to understand the mathematical causes for LLM failures, and therefore also possible to develop alternative technologies that overcome such barriers.

Our gut microbiome could be a driver of aging. Gut bacteria are generally considered to be a benefit to our health. Yet research described in New Scientist suggests those same bacteria might also accelerate aging, casting the role of the gut microbiome in a more complicated light. Up to the age of about 50, the new research suggests, the gut microbiome behaves pretty consistently. But as we get older, the diversity and function of the gut microbiome declines — a situation known as dysbiosis. Prior to that tipping point, beneficial bacteria outnumber more problematic ones; afterward, more aggressive bacteria can gain the advantage. This flip was thought to be a symptom of aging, but new research suggests that the shift also contributes to aging. The more problematic microbes are able to breach the confines of the gut, enter the bloodstream and cause mild inflammation throughout the entire body. This puts stress on the body and contributes to aging. Some researchers have come to believe that the entire microbiome, not just the bad microbes, is to blame for aging. And at least one study backs that up. In experiments, mice that were genetically engineered not to have a microbiome had a 26 percent longer life expectancy compared to a control group. To be clear, we can’t survive as a species without a gut microbiome: It is necessary for digestion and is the first defense against other, more dangerous bacteria that make their way into our digestive systems. And evidence is still accumulating about how, exactly, the microbiome is linked to aging. But the findings do suggest that attempting to preserve a youthful gut microbiome for longer — through exercise, a good diet, avoiding antibiotics as much as possible and even spending time with younger people — could help stave off the effects of getting older.

A flooded quarry, a mysterious millionaire and the dream of a new Atlantis, from The Guardian 
2,800 words, or about 11 minutes

Thanks to an anonymous and generous donor, the future of undersea life is being tested in an unlikely location: the idyllic countryside of the England-Wales border in the U.K. The donor has backed a company called Deep to design, build and test the infrastructure required to build human settlements on the ocean floor that will be used to study aquatic life and activity for long periods. Deep’s initial campus in the U.K. is costing $125 million, but this story claims that there is more than that amount of money on the table. The hope is to build modular accommodation units, known as sentinels, that could be lowered to depths of 600 feet or more below the ocean’s surface, potentially even linking up to create whole villages beneath the waves to establish a permanent human presence in the water as soon as 2027. For now, though, testing — of submersibles, accommodation units, and even banal realities such as catering — is taking place in a British quarry lake, and this story and its accompanying art provide a fascinating glimpse into what it looks like.

How a Leftist Activist Group Helped Torpedo a Psychedelic Therapy, from The New York Times
2,800 words, or about 12 minutes

Drugs like MDMA, psilocybin and LSD, which were once considered to be purely recreational, have all been studied extensively in small, preclinical trials and shown to help patients with post-traumatic stress disorder, substance use disorders, depression and more. But the future of psychedelics as legitimate treatments was set back last year, when the FDA declined to approve an MDMA treatment for PTSD developed by Lykos Therapeutics. The FDA's decision came as a surprise to some proponents, though many experts were sympathetic to the argument that more research was justified. Yet this story from The New York Times reports that the verdict may have been shaped by more than just a need for more scientific evidence. It describes how a leftist advocacy group called Psymposia, whose members — psychedelic advocates who oppose the commercialization of such drugs — waged a smear campaign against Lykos and barraged an FDA advisory panel with questionable accusations against the company — all the while passing themselves off as experts in the field of psychedelics, despite having no formal qualifications. The FDA ruling was not a death knell: The agency requested an additional clinical trial to further study the therapy’s safety and efficacy, and Lykos is still working with the FDA to gain approval. In the meantime the company has laid off 75 percent of its staff, and the verdict has muddied the conversation around a potentially promising field of medical treatment.

Motor neuron diseases took their voices. AI is bringing them back. from MIT Technology Review
2,500 words, or about 10 minutes

AI voice cloning can be put to use in some alarming ways, including for fraud and identity theft. But MIT Technology Review reports on a bright spot: bringing back the voices of those who have lost them to illnesses such as motor neuron diseases. Historically, those who suffered from such conditions could, if they had the stamina, “bank” a crude version of their voice by recording long lists of words. Now AI cloning means that a synthetic version of someone’s voice can be created from a short voice sample. (Though a voice replication can be created from as little as a minutelong recording, the longer the sample the better.) The results sound far better than the old “banked” voices, though they aren’t perfect. Users say, for instance, that the current voices lack emotion. The systems also require the text to be typed somehow, either with a phone, keyboard, or eye-tracking software, so conversations are slow. But if this article makes one thing clear, it is that regaining the ability to speak with one’s own voice — even if it isn’t perfect and it’s laborious to use — can be an incredibly emotional experience, for both the person and their loved ones.