Dear Aventine Readers,

Robots built like humans offer enormous potential for automating significant portions of factory and warehouse work, but for years have been plagued by clumsiness, poor balance and maddeningly slow speed. It is not so easy, it turns out, to replicate the human body’s unique combination of strength, balance and flexibility. This month we look at a new generation of humanoid robots that could change all that, with considerable excitement and funding behind them.

Also in this issue: Five experts weigh in on the state of AI-enhanced drug discovery, in which artificial intelligence is being used to speed up the process of identifying promising chemical reactions that could lead to new, possibly life-changing drugs. And don’t miss: a promising new way to make clean steel and AI systems that are being trained to learn language the way babies do.

Happy end of February and thanks for reading!

Danielle Mattoon
Executive Director, Aventine

Subscribe

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

Humanoid robots finally seem to be stepping up. At the end of last year, Amazon and GXO Logistics both announced that they were testing robots built by Agility Robotics in their warehouses. In mid-January, BMW announced it would try out humanoid robots built by Figure in its factories. And just a week later, Bill Gates wrote a blog post explaining why he was newly excited about robotics technology, singling out several humanoid robots projects. Tesla is even trying to build its own humanoid robot, Optimus, to help fully automate its production lines. Roughly $2.5 billion has been invested in the technology since 2021, according to PitchBook data, and Figure is reported to be in the process of raising a further $675 million.

This marks a striking turnaround from less than a decade ago. In 2015, with the goal of kickstarting a transformative age of human-like machines, the Pentagon’s Defense Advanced Research Projects Agency (DARPA) challenged robotics labs around the world to a contest. The machines would compete in performing numerous tasks like using power tools, opening doors and walking over rubble in the hopes of winning both a lot of attention and a share of a $3.5 million prize pot. But the competition was a flop. The robots were slow, awkward and prone to falling, inspiring not the hoped-for sizzle reel of innovation but bloopers that often went viral for the wrong reason. Instead of propelling the industry forward, the DARPA contest “totally killed it” for several years, said Dennis Hong, a professor at the University of California, Los Angeles, and founder of its robotics and mechanisms laboratory (RoMeLa).

So what’s behind the turnaround we’re seeing today? How “human” have these robots actually become and what are their limitations? To help answer those questions, Aventine spoke with experts in the field who explained that the technology has certainly turned a corner, but there’s still a long road of testing and development ahead before it becomes commercial reality.

“It is an exciting time, it's a great time to be in this field,” said Hong. “[But] I've been in this field for more than two decades. I know how difficult this is.”

There’s a Reason Robots Look Like Us 

It might seem like nothing more than narcissism that has led to so many robots resembling humans, but there are compelling economic and technical reasons to make robots in our image. 

“Our environments were built [along] with evolution over hundreds of thousands of years,” said Robert Riener, a professor at ETH Zurich, who recently published a paper in the journal Frontiers in Robotics and AI on the performance of robots compared to humans. Doorways and stairs are designed to be navigated by human-sized entities; tools fit in our hands and workbenches stand at waist height. Almost all of the world’s infrastructure is human-centric and that cascades through the economy, from domestic kitchens to commercial distribution centers. “If you don't want to change our environments too much, then we end up with humanoid structures [for robots],” Riener added.

The point is “not to build a robot that looks like a person,” said Jonathan Hurst, co-founder and chief robot officer of Agility Robotics and former professor of robotics at Oregon State University. “The goal is to do useful things in human spaces.” As a result, many researchers, including Hurst, prefer to use the term human-centered to describe the robots they design, arguing that they are building machines to work effectively in the human environment and with humans themselves — and that doesn’t necessarily mean they need to look like humans.

Some features might not look human at all: Agility’s Digit robot, for instance, has legs that wound up resembling those of a bird because, the company argues, the physics of bipedal locomotion show that such a design provides the most stable robotic legs. Still, many things that you might ask a robot to do in a human-centered world work better when a robot has a human-like form. An upright torso helps a machine to balance, for instance, while arms that are attached at the top of the torso make most sense for bimanual manipulation of objects and reaching out to take the impact of a fall. As for robot heads? They usually serve no purpose other than to make humans feel more at ease. 

What’s changed?

In short: a lot. “There's no one thing I can point to and say, ‘Hey, this is what's making it better,’” said Hurst. Instead, ask robotics experts and they’ll give you a long list of advances that have collectively contributed in some way to the recent uptick in performance of humanoid robotics. 

Take walking. Robots have improved their walking skills so much that Hong said many people believe that it’s almost at the point of being a “solved problem,” meaning no more stumbles and falls, or at least far fewer. To Hurst’s point, this leap forward isn’t the result of a single advance, but of many. Increased understanding of the biomechanics of walking in animals, better actuators that behave more like biological muscles, improved systems for controlling those actuators and advances in artificial intelligence have all played a role in this improvement, experts told Aventine.

There are many other advances that have helped improve general performance too, and Riener reeled off a list: Better batteries mean that robots can now carry enough power to be useful for hours rather than minutes; less expensive and better quality sensors allow robots to navigate their environments better; new motors allow them to perform slow and fast movements with equal competence.

One area where there’s still room for improvement among all robots, however, is in the manipulation of objects — a topic that Aventine reported on in depth in 2021. On this front, the human hand, which is highly flexible, is proving to be an almost impossible goal for robots. Advances in areas such as computer vision and gripper technology have resulted in robots that can manipulate objects more effectively, and particularly at highly specific tasks, but robotic manipulation lags “far behind average human skills,” Riener wrote in his recent paper. 

All told, the surge in humanoid robot competence is due to a confluence of technological improvements, though the results haven’t been even. A state-of-the-art robot may now be good at walking, but it’s better able to perform a simple task like picking up a box than dicing carrots for your dinner.

How useful are they, really?

Because humanoid robots are not yet in commercial use, to see what they’re capable of we often have to rely on PR demonstrations, which by their nature deliver a rosy view. Tesla’s Optimus is able to neatly (though slowly) fold a T-shirt, but seemingly only when controlled by a human. Figure’s 01 robot, according to the company, taught itself to use a Keurig coffee machine and was then able to make a cup of coffee, though again, far less quickly than a human could. Perhaps more prosaically, Agility’s Digit robot can move plastic warehouse crates — its only use for now — without looking awkward and jerky. Compared to those 2015 DARPA fails, it is fluid and impressively utilitarian, exemplifying the kind of skills that might be closer to prime time. 

All three companies, as well as others like Apptronik and 1X, see that despite increased automation over the past few decades, large sectors of the economy still rely on humans doing repetitive tasks. E-commerce warehouses, for example, despite being highly automated, require humans to move large containers from one automated system to another. Similarly repetitive tasks are a reality in other factories and warehouses around the world, and the well-reported U.S. labor shortage means that many companies are hungry for the help automation can provide — even if that is, for now, just moving boxes.

“You have to start with something,” said Riener. “Maybe the best in the world is [currently] very simple.”

The challenges ahead

While recent strides are significant, plenty of work remains to be done. “Just because your robot can walk and do things does not mean you can use it in a factory in a meaningful capacity,” said Hong.

Much as no single advance can account for recent improvements in humanoid robot technology, no single advance will lead to their future improvement. “There's nothing fundamentally that is like, ‘Hey, we need to invent this silver bullet and then all of a sudden, it'll [all] be possible,’” said Hurst. Hong, meanwhile, when asked what advances were required to keep pushing the field forward, replied simply: “Everything.”

Still, researchers and companies seem to have their own priorities based on what they’re interested in achieving. For Riener, advanced control systems that will be able to parse inputs from myriad sensors and control increasingly complex hardware is an important area of focus; Hong suggested that more effort should be put into making the machines more resilient to failure; Hurst said that figuring out how to make the best use of new tools like generative AI could enable more sophisticated planning by robots, helping them patch together smaller processes to achieve more complex goals. And there are non-technical problems to solve, too, with Hurst pointing to regulation as an important factor in the process of rolling out these machines at scale in the future.

For now, the field is riding a wave — and like any hype cycle, that attracts interest and investment. “Now is when investors start to pour dollars in, [when] DARPA starts to pour dollars in, you know, everybody starts to want to pursue the dream,” said Hurst. “If there's one silver bullet [for progress], I can say it is resources: people and dollars.”

Yet all this comes with pressure to deliver. For some members of the community, the DARPA challenge lingers as a painful memory and a reminder of what could still happen if results don’t marry up with expectations. The nascent industry could take off, flush with cash and technical advances — or it could flounder. “I don't have a crystal ball,” said Hong. “I don't have the answer.”

AI that learns language like a baby. Researchers from New York University have carried out a study using video and audio recordings taken from the perspective of a young child to train an artificial intelligence system. This approach marks a sharp contrast to the prevailing method of training large language models on data sets containing billions of words, and also suggests a possible path toward training machines to learn in a more human-like way. The research, published in Science, used recordings captured by a camera worn by a baby for 1 percent of its waking life, between 6 months and 2 years of age. The sixty-one hours of footage, broken down into still images and 250,000 transcribed words uttered by humans during the filming, were used to train a neural network. When asked to match a word to one of four images, the software was correct 62 percent of the time — comparable to a conventional computer vision algorithm trained on 400 million image-text pairs, but not as impressive as a typical 2-year-old. Perhaps most important, though, the finding challenges the idea that the human brain is specially wired to understand language and that general AI algorithms will forever struggle to achieve human-style language learning. Next up for the researchers, MIT Technology Review reports, is to investigate how they could tweak their algorithms — to, say, include more contextual information, such as what parents are looking at when they talk, or basic physics models to replicate children’s intuitive understanding of the physical world — to improve the model’s chances of keeping up with a 2-year-old.

A cleaner way to make steel. A new technique that uses electricity and salt water to create iron hints at a future in which steel can be made without the huge levels of CO2 emissions that currently plague its production. The process, described in the journal Joule, uses a battery-like setup to strip oxygen atoms from iron oxide to create pure metallic iron, with chlorine and sodium hydroxide as byproducts. The researchers claim that the process is efficient enough that the sale of the chlorine and sodium hydroxide — chemicals that are both used industrially — could cover the cost of iron production, while renewable energy sources could power the reaction to ensure that it was carbon-neutral. This would be an improvement over the current approach to extracting iron from its ore, in which it’s heated to 1,500 degrees Celsius with coal, contributing to about 8 percent of all global CO2 emissions. As ever, there are caveats, principally that the technique hasn’t been scaled up yet, and doing so would require an efficient method for sourcing clean iron oxide as well as finding markets for all those byproducts. But the achievement is nevertheless a precursor to a possible future in which steel can be manufactured and not add to carbon emissions.

The engineering marvel — and questionable promise — of Apple’s Vision Pro. It is nine months since Apple unveiled its VR headset and the reviews are finally in. The quick take: It’s an imperfect yet spectacular feat of engineering and a hint of what the future may hold. The Verge’s Nilay Patel goes deep into the weeds with his review, and the detail is fascinating: The device’s two tiny screens, for instance, contain more than 23 million pixels, each about the size of a red blood cell. More broadly, Patel is impressed by the device’s unrivaled ability to transmit video from its surroundings into the eyes of the wearer so they experience real and virtual worlds at once. As a consumer device, though, it’s flawed: heavy, expensive, and often frustrating to use. What happens next is up for debate. Longtime tech analyst Benedict Evans wonders if the device and others like it are destined to be niche products. But the reflections of Vanity Fair’s Nick Bilton inspire the biggest pause for thought: “I can see a day when we all can’t imagine living without an augmented reality,” Bilton writes. “When we’re enveloped more and more by technology, to the point that we crave these glasses like a drug, like we crave our iPhones today but with more desire for the dopamine hit this resolution of AR can deliver.” 

OpenAI built a powerful text-to-video AI. What happens when the abilities of ChatGPT and DALL-E collide? Impressive artificial video and some potentially huge ethical headaches, as we found out this month with OpenAI’s release of Sora, a generative video model that converts short text descriptions into highly detailed videos. The technology uses the abilities of diffusion models, such as those used by DALL-E, to create high quality video frames, while transformer models, such as those used by ChatGPT and other large language models, are employed to put images into sequences. There are caveats: The videos have clearly been cherry-picked by OpenAI to tease the technology, and even then there are plenty of errors — from spontaneously appearing humans to physically impossible images, such as candle flames burning at unnatural angles. Yet the ability to create video through generative AI has existed only since 2022, and in that time the quality of the video it can produce has gone from grainy and glitchy to something that could easily be mistaken for genuine. That raises substantial ethical concerns, principally around the technology’s susceptibility to be used to create ultra-realistic fake video. OpenAI says it isn’t releasing the tool for general use yet over concerns about potential misuse, and is instead evaluating it with third-party safety testers and video makers. Yet make no mistake: That testing surely means that OpenAI is anticipating commercial release of this technology at some point, and there is little to no chance it will be used solely for good or benign purposes.

Discovering and creating new drugs is a notoriously slow and difficult process. The preclinical stages of development — discovery, creation, and early testing — can take anywhere from three to six years. This investment in time also makes drugs expensive, as the pricing for a drug is often tied to the amount of money that went into its creation, ranging from hundreds of millions to even billions of dollars in recent years. 

Artificial intelligence models could potentially transform this development cycle, speeding up the process and reducing the cost. AI systems trained on chemical or biological data sets or information about specific diseases could create lists of compounds or molecules that might target diseases in unexpected ways. They could even design molecules that don’t currently exist.

Several companies have already introduced drugs into human clinical trials that were developed with AI assistance, including drugs for pulmonary fibrosis and cerebral cavernous malformation. Such successes, paired with fever-pitch hype around AI, has prompted large drug developers and manufacturers, small technology companies and AI researchers to pour time and money into AI-enhanced drug discovery.

Experts in AI and drug development all told Aventine that they are confident that these systems will eventually become sophisticated enough to seriously accelerate drug discovery. But they also called attention to several challenges in mixing artificial intelligence with medicine, like the knowledge gap between technology experts and the chemical, biological and medical researchers responsible for developing drugs. Additionally, the AI models built for this work share the problems that plague all generative AI systems, like Open AI’s ChatGPT and Google’s Gemini: The models are not aware of what they produce and lack the ability to assess the quality of their results. This means that a generative AI system programmed to suggest drug models could easily produce hundreds if not thousands of new but useless chemical reactions, requiring a long cycle of testing before human researchers find a potentially useful molecule or drug. As a testament to how difficult it is to discover new drugs, however, this process is often an improvement on the needle-in-a-haystack alternative, as researchers say that the AI-generated results are usually at least directionally helpful. Ultimately, those in the field believe AI models can be trained to be more targeted and efficient by increasing the quality of the data the models are trained on, though those data sets can be hard to find and procure and expensive to create from scratch. Finally, several experts also cited concerns over the threat that the approach could be abused to design poisons, toxins or chemical weapons.

1. California, U.S. There’s a technique known as forensic DNA phenotyping, a method of predicting some of a person’s physical traits through a DNA sample, and then creating an artistic impression of what that person might look like. The technique, which can accurately predict only general characteristics like skin and hair color and facial shape, is intended to be used by law enforcement agencies to narrow down pools of suspects and unidentified crime victims. But Wired reports that at least one police department in California has attempted to put one of these generalized images through AI facial recognition software in an attempt to identify the perpetrator of a murder. The problem, according to facial recognition experts, is that phenotyping is an inexact science and the results are merely an indication of certain features a person may have, not an accurate rendering of what they look like. While facial recognition software is now highly accurate when used on good quality photos, it is far less so when it is fed low quality images and illustrations. So, as one expert told Wired, “daisy chaining unreliable or imprecise black-box tools together is simply going to produce unreliable results.” Perhaps even more important is that there’s currently no federal oversight of these tools — meaning that, for now at least, police departments and technology providers get to determine if this kind of use is acceptable.

2. Bangladesh. While the U.S.might be the birthplace of modern software development, other nations are quickly expanding their capabilities. Rest of World studied data published by the code repository platform GitHub to evaluate an increase in the number of developers around the world, and found that the number of accounts in Bangladesh rose by 66 percent in a single year, from 568,145 developers in 2022 to 945,696 in 2023. The second-highest growth was seen in Nigeria, at 45.6 percent. What’s less clear is the exact meaning of such numbers. While they could be an early indicator that a nation’s fledgling tech industry is taking off, they could equally be evidence of a nation learning to code but without access to funding and jobs to take advantage of those skills. Undeniable, though, is that software development now reaches far and wide — into almost every economy on the planet.

3. Everywhere. Or at least in one million locations around Earth. That’s the number of places that could technically be suitable for pumped storage hydropower facilities — systems that pump water from low reservoirs to higher ones while electricity is cheap and plentiful and then release the water during times of demand to drive energy-producing turbines — according to research by a team from the Australian National University. To be sure, such systems have existed for decades, but Science reports that, as we face up to the reality of periods when solar and wind facilities can’t keep up with demand, we will need more of them. Many such facilities are currently being built and are going online in the near future — particularly in China, where 66 new plants are reported to be under construction. Yet in many locations, debate rages about whether these plants, which are hugely expensive and take years to build, deserve such a prominent role in our energy future. Is the infrastructure too destructive of nature? Are we confident enough in our projections of demand over the coming years to warrant them?

The Case for Nuclear Cargo Ships, from IEEE Spectrum
3,000 words, or 13 minutes

Shipping accounts for about 3 percent of the world’s greenhouse gas emissions. There are plenty of proposed alternatives for the diesel that currently fuels the majority of container ships: Liquefied natural gas and methanol are already advanced as alternatives, with hydrogen, ammonia and even batteries being considered too. This article takes a close look at how small modular nuclear reactors could be used to power ships and enable them to go years without refueling. Though the new small modular reactors are reportedly safer to operate than conventional reactors, the technology is still dogged by negative public perception and remains unproven at sea. That said, if planned tests of SMR-powered ships are successful, shipping could become greener far faster than we hoped.

Ovaries Are an Enigma That Could Unlock Human Lifespan, from the Financial Times
4,100 words, or 18 minutes

Centuries of sexism have given rise to a peculiar situation that persists today: Although pregnancy and childbirth are central to human existence, research science around them has been neglected, meaning that they remain relatively poorly understood. Nicole Shanahan, an entrepreneur turned philanthropist from California, has been trying to change that. In 2019, she committed to spending $100 million to help fill in much of the missing knowledge around women’s reproductive health. This story from the Financial TImes takes a close look at her motivations, the work she’s funding — including experiments on ants and naked mole rats — and how the research could transform reproductive longevity, potentially making it easier for women to choose when they have children.

The New Car Batteries That Could Power the Electric Vehicle Revolution, from Nature
3,500 words, or 15 minutes

Building car batteries is a tough gig. Ideally they should be light, high-capacity, high-powered, fast to charge, slow to degrade, efficient across a wide range of temperatures, safe … and, oh, add cheap to the list too. For now, lithium-ion batteries are the only option, but with so many parameters to optimize for, it’s perhaps not surprising that researchers around the world are trying out all sorts of new approaches to build better alternatives — or even just batteries that are better against some of those metrics listed above. This story from Nature takes a close look at the competing technologies, and finds that some of them may be inside our cars sooner than you might expect — perhaps even next year.