August was a nice reprieve from travel. I went hiking, cleaned out my shed, and fended demon squirrels from the nine-foot-tall sunflowers in my backyard. But here we are, at the cusp of new season, a slight chill in the air and candy corn by the Walgreen’s checkout, signaling the arrival of a new school season, and you in your autumn sweater.
I, meanwhile, will be back on the road. Those in the Steel City of Bridges on September 16 can catch me hosting a fireside with Seegrid CEO Chris Baker on the Main Stage of Pittsburgh Robotics Network's Robotics & AI Demo Day.
If you're in the Bay Area and remember the 22nd night of September, I'd love to see you at Automated's first-ever West Coast Live Podcast and Happy Hour at Circuit Launch in Mountain View. We'll kick things off with a taping of an Automated episode with Robust.AI CTO and all-around robotics legend Rodney Brooks, and transition into food and drink-fueled schmoozing.
I also hope to see some of you at A3's Advanced Vision and AI event on September 23-24 in Santa Clara. Past Automated guest and current Waymo distinguished engineer Vincent Vanhoucke will be there, among others, while I will be hosting one of the stages both days.
Until then, we've got a packed newsletter below. Google DeepMind's head of robotics, Carolina Parada, discusses her journey and the decisions that shaped Gemini Robotics. Also, Jagdeep Singh returns to talk about Rhoda AI's hardware ambitions.
We've got Chef Robotics CEO Rajat Bhageria on the pod, while Becca profiles Perceptual Robotics, and we finally let Liam write about soccer. Puns are made.
"Faced with information overload,” Marshall McLuhan wrote in 1969, “we have no alternative but pattern-recognition."
Patterns are how we find signal within the universe’s overwhelming noise. The ability to recognize them is among the world’s most transferable skills. It’s no doubt one of the reasons so many professionals have made successful transitions to robotics from other fields. Just the other day, we spoke to a physical AI founder for a forthcoming interview with a background in biochemistry and molecular biology.
While Carolina Parada’s own CV doesn’t present quite the same level of whiplash, it does illustrate a kind of non-linear throughline. Before stepping into her current role in May, Google DeepMind’s VP and head of robotics transformed a deep background in language modeling, search, and speech into a principal deep learning engineer role on NVIDIA’s self-driving team.
In a recent conversation with Automated, she was quick to recognize a pattern connecting her own seemingly disparate focuses. “If you notice in each one of them, I was there trying to change the way the current community was solving the problem,” Parada explains.
“I think in most of these cases, I try to go at it from how can I learn this from data in a way that feels robust and resilient to noise,” the executive adds. “How do I try to solve this without necessarily evaluating against the current techniques, but not just trying to do an incremental improvement on them. It's perfectly okay to sort of start from scratch and remove the existing approach before evaluating against it.”
Tabula rasa is a terrifying prospect for most — particularly in a field to which you’ve devoted much of your professional career. It doesn’t mean ignoring the generations of advancements that have enabled the current methods, so much as challenging whether they remain the most effective ways of approaching a given problem.
It’s easy to imagine a world wherein such a widescale questioning of orthodoxies from a new team member might be cause for consternation. Parada tells me, however, that DeepMind Robotics was already operating under a similar mindset when she joined the team.
“You might be surprised to hear this,” says Rhoda AI CEO Jagdeep Singh, “but we think that we could be the first autonomous manipulation-capable robot that's deployed at scale.”
I am, in fact, surprised to hear it. Beating the dozens of industrial humanoids to market would certainly be a coup for a system that has yet to be announced. Since emerging from stealth in March, Rhoda’s outward-facing work has focused on its model training.
The Bay Area startup has showcased a novel approach to physical AI that relies almost exclusively on video for pre-training data. Public demos from the company, meanwhile, have featured off-the-shelf robotics hardware, showcasing the model’s adaptability to different form factors.
Singh says the company has been focused on building a full-stack hardware/software solution since its earliest days. Rhoda attempted developing for existing systems, but found them insufficient for the kinds of industrial tasks it was looking to address.
“There are probably over 100 humanoid companies around the world, but most of them, to be candid, are toys,” he says. “So, they can lift a small amount of weight, but not a lot. And in even that weight that they can supposedly lift, they can't do it day in, day out reliably.”
Rhoda hasn’t been secretive about its hardware plans, exactly. In previous conversations, Singh and I have discussed the company’s work in the humanoid space. Still, no official announcement has been made. When we touched base earlier this week, the CEO addressed the company’s staggered timeline.
“We very deliberately and very consciously don't want to show hardware or software that isn't ready for prime time,” says Singh. “It causes us to be a little later in what we announce, but what we do announce, we want it to be real. We don't want to be one of the many, many companies in the space that are showing demonstrations and not deployable product. Everything we've talked about so far has generally been things that we, you know, have a high level of confidence in relative to actually being deployable.”
Restaurants only operate at peak production for a few hours a day. The economics and technical reality did not support the company Rajat Bhageria wanted to build. Watch on YouTube>
Kate Darling (RAI Institute) - Is emotional attachment to robots actually a good thing? Can we give human workers a stronger voice in the conversation around workplace automation? These big picture questions and more.
Nic Radford (Persona AI)- A lot of companies claim to be building humanoids for the real world, but few walk the bipedal walk quite like Persona AI's ultra-rugged welding systems.
Kathryn Zealand (Skip) / Jessica Bath (UCSF) - Kathryn Zealand joins us to discuss her Google wearable robotics spinoff, Skip. After that, UCSF assistant professor Jessica Bath discusses how exoskeletons are being used by Parkinson's patients.
I would like to personally thank NVIDIA and Hugging Face for officially confirming the much discusses $12.9 billion acquisition deal just as we were putting the final touches on this morning's newsletter. Really appreciate your letting us sneak this in there before publishing. "Hugging Face will remain an open platform for the entire AI ecosystem. Developers will choose the models they want, the frameworks they want, the clouds and inference service providers they want, and the computing platforms they want," NVIDIA's Jensen Huang said in a post. "NVIDIA compute will not be required to build on or deploy through Hugging Face. Hugging Face will continue to support open-source and open weight models from across the ecosystem, from every model builder. It will continue to support multi-cloud and multi-accelerator development and deployment, so builders can use the hardware and infrastructure that best fit their work."
Hugging Face's considerably wealthier CEO, Clem Delangue, seemed pleased with the news in a note posted to LinkedIn, stating, "Ten years after starting Hugging Face, open-source AI is at an inflection point. Thanks to the community, we’ve shown that it can be a complement, and even an alternative, to closed-source APIs. But for it to happen at larger scale, it needs more compute, more support, more collaboration, and more visibility. That’s why we went to talk to Jensen, who offered to do exactly that with us. In addition to doubling down on NVIDIA’s massive contributions to open-source AI (I called them the 'King of American open-source AI' earlier this year), they’ve committed to strongly supporting Hugging Face and our mission while keeping the platform open, independent, and compute agnostic. The founders and the team are all staying to keep pushing this mission forward."
Since emerging from stealth, World Labs has been among the spatial AI startups for one simple reason: Fei-Fei Li. The long-tenured Stanford professor is considered AI’s Godmother, thanks in no small part to her foundational visual database, ImageNet. World Labs hit the scene in late 2024, with $230 million in funding — small potatoes compared to the $1 billion the firm raised earlier this year, thanks to big names like NVIDIA, AMD, and Autodesk. This week, the company announced its new world model, Atlas, which will power its 3D world modeling program Marble and other offerings going forward. “All inputs are combined into a shared spatial context,” the company writes in the announcement post. “Atlas uses that context to generate what comes next, staying consistent in 3D with everything it has seen and imagining what lies beyond it. Atlas is built to scale: its performance improves with increased training compute, and we expect this trend to hold as we continue scaling.Atlas can perform a broad range of tasks spanning world generation, reconstruction, and simulation.”
Call it the "Better Late Than Never Fund." Andreessen Horowitz ("a16z" to my down-vesting Valley friends) recently announced the launch of its $1.1 billion "Machine Age Fund," with a graphic that screams "Ayn Rand giving Megalopolis a script treatment." The fund finds the firm making hardware (and, by extension, robotics) an "official motion" in a bid to "open the throttle and accelerate the physical buildout of AI: the strongest tool ever developed for solving problems and bestowing abundance." While it hasn't traditionally been a hardware firm, a16z points to some success stories over the years, including Skydio and Waymo. More recently, it's boarded the physical AI rocket ship courtesy of investments in Unconventional AI, Nexthop, Volta, Atoms, Heron Power, and Mind Robotics.
CES 2026 darling Lyte just announced a $165 million Series C valuing the robot sensor firm at $1.6 billion, post-money. Lyte was founded in Silicon Valley by engineers from Apple and PrimeSense (the Apple-acquired startup behind the Xbox Kinect camera). The company developed its own perception stack and has already started shipping product to industrial robotics clients. “Physical AI has a sensing problem before it has a model problem,” says CEO Alexander Shpunt. “A robot cannot act safely on data that does not faithfully describe the world. We build the entire perception foundation, from custom silicon to the spatial data our systems produce. This financing will put [perception platform] LyteVision into more robots, across more environments, faster."
Inspection has long been one of the drone world’s most valuable applications. They’re small, relatively inexpensive, carry cameras, and can safely reach spots we’d rather not send humans. Remote energy facilities like nuclear reactors, oil rigs, and wind farms have all benefited from the technology’s proliferation, allowing important inspections to occur on a far more routine basis. This week, Becca profiled Perceptual Robotics, a Bristol, U.K.-based firm that build drones specifically for wind turbines. The firm is utilizing AI to find proverbial needles in haystacks. “The blades are sometimes up to 100 meters or more, and then you're trying to find something that might be 3-4 millimeters big on that 100-meter blade,” CEO Kostas Karachalios told Automated. “You need very high quality, huge volume of images to be able to achieve that. And this is on something that normally looks like a plain white or gray surface, right? So, it's very difficult to find features on this kind of imagery. It is a hard computer vision problem to solve.”
Judging from the headline, "Is Adrien Robot the Next Adrien Rabiot?" Liam had entirely too much fun writing this one — in spite of the fact that our American sensibilities forced him to refer to his favorite pastime by the s-word. The piece covers research that expands the concept of footie-playing humanoids beyond the RoboCup. "We wanted to explore whether reinforcement learning could instead learn this perception-and-action loop directly," lead author Yushi Wang told Automated. "Rather than treating vision, planning, locomotion, and kicking as completely separate modules, we wanted the robot to learn how to react to visual information through its own motion. This led us to develop a unified vision-driven controller that could search for the ball, approach it, adjust its gait and body orientation, and kick — all as part of a continuous reactive behavior."
FieldAI continues to rack up high-profile partnerships. This week, its construction giant Caterpillar, which will utilize the Pittsburgh startup’s physical AI tech (and, judging by the images, a lot of Spot robots) to perform autonomous safety inspections, create jobsite digital twins, and improve risk assessment. Here's FieldAI CEO Ali Agha: "Caterpillar has shaped operations on job sites and manufacturing facilities for over a century, and these are exactly the environments FieldAI excels in, with deployments spanning hundreds of sites worldwide. This collaboration brings leading physical AI capabilities to a leading manufacturer of construction and mining equipment."
Researchers at MIT this week showcased reconfigurable, modular blocks that maintain electrical connections as they change shape. The implementation of these “bifur-circuits” allows for a broad range of functional applications through bendable conductive material that can be compressed more than 10,000 times. “Metamaterials can make complex mechanical assemblies easy to manufacture just by using repeating units,” says lead author Marwa AlAlawi. “Our work expands on this design space. If we think of mechanical metamaterials as building blocks, then our work is one way to take advantage of their geometry to embed intrinsic intelligence into hardware, which could open many possibilities.”
It was a bit tongue-in-cheek positioning Microduck as Hugging Face's "big news" last week, in light of ongoing reporting around an NVIDIA acquisition. Still, the French AI firm's robotics subsidiary, Pollen, has a bona fide hit on its hands. The little system has sold more than 10,000 units in its first five days. An impressively webbed feat for a $400 open-source robot.
Wake up, babe. A new Y Combinator startup just dropped. Robocurve is building "benchmarks for robots and the AI models that control them." This includes all of your standard robot stuff: sandwich making, Jenga playing, data center building. You know the drill. "General-purpose robots could arrive before the end of the decade, with far-reaching implications: accelerated data center buildouts, explosive economic growth from a fully automated economy, and intensified geopolitical competition between the U.S. and China," writes cofounder Jay Chooi. "Real-world evaluations from Robocurve will help clarify the timeline to general-purpose robots so that society can prepare for their arrival well ahead of time."
The Association for Advancing Automation (A3) is North America’s largest automation trade association representing more than 1,400 organizations involved in robotics, artificial intelligence, machine vision & imaging, motion control & motors, and related automation technologies.
.png?upscale=true&width=1120&upscale=true&name=Rajat%20Bhageria%20(1).png)
