Greetings from Nashville Airport’s Terminal B. Over the past 24 hours, the time I haven’t spent here has been equally split between the hotel and a massive warehouse that my Uber driver told me was a big, empty field until fairly recently. In fact, a couple of hours before writing this, Dexory’s three cofounders huddled around a single pair of scissors to cut a symbolic — yet literal — ribbon to mark its grand opening.
It’s got that new warehouse smell, which is to say, it doesn’t smell like anything in particular. There’s row after row of shelving units packed with big, brown boxes. The boxes, on the other hand, contain nothing. I nudged one with an elbow as I passed by, just to be certain.
It’s a fun, full-circle moment for the London-based logistics automation firm. Before kickoff, CEO Andrei Danescu pulled out his phone and opened the Photos app. The lead image was a picture of Dexory’s first robot (which only stands a modest 40 feet tall to the new version’s 60 feet) parked inside a warehouse not dissimilar to the one in which we currently stood.
He was up early this morning, he said, when the app surfaced the image as a memory. The shot was taken as the company was delivering the system to its first client — three years ago this week. The timing, he assured me, was purely coincidental.
More info on Danescu, Dexory, Formula 1, and more, when our full podcast interview drops in a few weeks. This week, however, it’s a chat with another startup that has carved out its own successful niche over the last few years. Matic’s president and cofounder, Mehul Nariyawala, talks about building a better vacuum, the market following iRobot’s Chapter 11, and what form a home robotics platform might actually take.
Much of our recent coverage has focused on what we’ve informally termed the “liminal zone” of robot training. While there’s a consensus that real-world data collection is a critical piece of developing robust physical AI models, it’s unclear how we can deploy enough robots to collect that data at scale without the aid of such models. In other words, how do we kickstart the flywheel?
In an interviewe tied to Ai2’s recent MolmoBot launch, the organization’s director of Perceptual Reasoning and Interaction Research, Ranjay Krishna, noted that — while there are plenty of industry robots currently deployed in the real world — many are fixed and tasked with highly repetitive jobs. These systems aren’t especially useful when it comes to building dynamic world models.
“What we really need is robots working in unstructured environments for this data flywheel to become useful — places and environments where it's not clear what's going to happen in those environments and your robot has to adjust its behavior,” says Krishna. “In your home, maybe you spilled something, maybe you fell down, you'd want your robot to do something in response to things that you're doing. We don't even have robots good enough or safe enough even that we can deploy them in these unstructured environments. The state of Flywheel, to get it even started, we need robots to be good enough where we can put them out in some unstructured environments and we're just so far away from that.”
How difficult and expensive will it be to collect real-world data at scale? Opinions vary. Physical Intelligence cofounder, Sergey Levine, recently told me that many in this space are overestimating such challenges.
“In the grand scheme of things, maybe sometimes something that we as researchers get a little bit mixed up about is easy for us versus easy in the context of human civilization,” he said. “And in the context of the world as a whole, actually getting real-world data and then deploying robots that are going to collect more experience and get better and better is a lot easier than inventing some other technologies just to avoid having to do that. And because it's a bootstrap problem, it's easier once things are out in the world.”
In our conversation, Krishna addresses some of the key challenges facing this real-world data gathering.
“Whenever we've had to go out and collect data ourselves, it's not just that you need the hardware and everything set up correctly, but you need to train people to give you data that's actually useful for training robots,” he notes. “And the training process isn't easy. You have to get people to move in ways that feel unnatural to them. They're usually tele-operating these robots. They have to learn the form factor of the robot that they're moving. It's quite a lot of actual setup that goes into the process before you even have a usable data collection process.”
“People ask, what's your biggest difference?” Wei Ding says with a smile. “I think it’s the shoes.” The earliest iteration were purchased at discount chain, Ross Dress for Less for around $20.
They lasted all of a month, forcing Noble Machines to upgrade Moby’s footwear to something more befitting a hulking metal humanoid: Caterpillar work boots. Tough robots deserve tough shoes, even at 5x the price.
The Bay Area robotics startup’s CEO sees something deeper in the decision to lace up the robot. Ding notes that company could charted a wholly different route, building the rubberized robot feet, but even in this world surrounded by billion dollar Silicon Valley startups, sometimes the simplest solution is the best.
"We want to use validated system,” he says. “For technology that's already proven and our partners or suppliers are better at it, we try to not reinvent the wheel. A unique part of us is we use off-the-shelf actuators.” It’s a philosophy you’ll find throughout much of the industry robotics market, wherein companies regularly opt for third-party arms, rather than building their own, in-house. In the world of humanoid, however, most companies seem determined to build as much of a proprietary hardware stack as humanly possible.
Chief among the advantages to this approach are a significant reduction in R&D costs and a substantially shorter time to market. Speed has been foundational to Noble Machine’s approach. The startup is determined to prove out go-to-market as quickly as possible.
That, in part, requires building robotics as quickly as possible. Founded in May of 2024, the startup publicly showcased its first robot at the Bay Area Humanoid Summit that December. “We built our third-generation humanoid from scratch with four co-founders in four months,” says Ding. “That was a pretty fast for us.”
Hands are — at once — humanoid robots’ greatest asset and biggest challenge. They form the foundation of much of what we imagine these systems accomplishing, both at work and in the home. Mobile manipulation, however is incredibly hard. For this reason, pilots have largely taken the form of baby steps — training and deploying systems to master one highly repeatable task with near-perfect accuracy.
“Most of our competitors are very, very focused on manipulation,” says CEO Arnaud Robert. “It has different flavors. It can move boxes around, triage objects, etc. We felt it was a natural entry, but the humanoid only did that, it would be quite limiting over the long run.”
Boiled down to a single word, Hexagon’s approach to the form factor can be described as: “sensors.” In three, it would be: “lots of sensors.” Humanoid manufacturers have largely prided themselves on their speedy hardware iterations — a description that certainly applies to the swift launch of Aeon — but Hexagon is hoping to stuff its robot with some added futureproofing.
“When you look at a more complicated task, as I mentioned, for scanning a door that might be 10 meters to 30 feet away for you, you need a lot of spatial intelligence,” says Robert. “But if a human actually crosses the path, what do you do? If there's a change in the position of the door, what do you do? And so on and so on. So when we look at all this, and again with the spirit of it has to be multi-purpose, we very quickly landed on, we need a pretty sophisticated suite of sensors if we want to be future proof from that perspective.”
There is tremendous opportunity in automating various manual labor jobs, from tote moving, on up. For Hexagon, however, “future proofing” means thinking about roles beyond material movement. Inspection is an immediate possibility for systems with sufficiently powerful sensors.
“When you look at what other things [a humanoid] could be doing, one of them is asset or part inspection,” says Robert. “The humanoid could do this because it can roam around a factory, find the part to inspect and inspect it. It can do space inspection if you have, whether for safety, for security, or for other purposes. You can do reality capture, and you can do more sophisticated things where the humanoid effectively is doing a task that requires taking an object, going to a door, scanning the door in high resolution.”
It’s features like parts inspection that made Hexagon stand apart from its humanoid brethren when BMW went searching for some next-generation automation assistance. There is, certainly, something appealing in a robot that doesn't simply transport parts, but rather inspects each for potential issues in the process.
I’ve been covering this industry long enough to remember when being valued at $1 billion was a big deal for a startup. The AI bubble boom laughs at your silly unicorns and sends them galloping back to the Lisa Frank backpack where they belong. If you want to raise serious money, get yourself a Yann Le Cun, a French knight who spent a dozen or so years building off Meta’s AI bona fides. Tuesday marked the long-awaited debut of his physical AI firm, Advanced Machine Intelligence (AMI) —pronounced like “friend” in LeCun’s native tongue.
The announcement included a $1.03 billion round, featuring Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expedition. Here’s a very long list of humans and organizations that have supported the Paris/New York/Montreal/Singapore-based firm: Toyota Ventures, New Legacy Ventures, Temasek, SBVA, NVIDIA, Mark Cuban, Association Familiale Mulliez, Groupe industriel Marcel Dassault, Sea, Alpha Intelligence Capital, Eric Schmidt, Aglaé Lab, ZEBOX Ventures, Artémis, Xavier Niel, Publicis Groupe, Samsung, Bpifrance Digital Venture, Jim Breyer, Tim & Rosemary Berners-Lee, and Mark Leslie.
“It is time to move beyond shortcuts and work on a foundational solution. World models learn abstract representations of real-world data, ignoring unpredictable details, and make predictions in representation space,” CEO, Alex LeBrun, wrote on LinkedIn. “Action-conditioned world models allow agentic systems to predict the consequences of their actions, and to plan action sequences to accomplish a task, subject to safety guardrails. AMI will advance AI research and develop applications where reliability, controllability, and safety really matter, especially for industrial process control, automation, wearable devices, robotics, healthcare, and beyond.”
For the first time in Automated’s six-month history, a $450 million Series A is a distant second funding round for the week. Call it getting “Yanned.” Pushing half-a-billion is obviously an unimaginable sum of money by practically any measure, of course, but that’s doubly the case for a company no one had heard of until this week. Given how things are going, you’ve no doubt already discerned that Rhoda is in the physical AI biz. Today the Bay Area firm announced FutureVision, a Direct Video Action (DVA) model for robot training.
“We believe the next era of robotics requires models that understand how the world moves — not just what it looks like or how it’s described in language,” said cofounder and CEO, Jagdeep Singh. “By learning from internet-scale video and operating in closed loop, our systems are designed to adapt to real-world variability in ways conventional approaches struggle to achieve. The goal is simple: robots that work in the real world, not just controlled lab settings.”
As for how DVA differs from the more familiar VLA, Rhoda notes the much-discussed issues of real-world variability and unstructured environments. The solution? First: lots of videos. Lots and lots of videos. Hundreds of millions, per the company. Robots are then post trained using a significantly smaller amount of data targeted at specific tasks. Says Rhoda, "The resulting system continuously observes its environment, predicts future states as video, converts those predictions into actions, executes them, and re-observes the world — repeating this process every few hundred milliseconds in a closed loop."
Hey, will you look at that? It's our first story from one of our new news freelancers. Shout out to Liam Critchley. He's a U.K.-based science writer, whose work has appeared in places like IEEE Spectrum and Interesting Engineering, so he'll be helping expand our research coverage. This week, he hasn't strayed too far from home, as the UK Atomic Energy Authority (UKAEA) and CERN have teamed to create ‘PipeINEER’ (Pipe+Pioneer), an autonomous race car for mice designed to inspect the inside of the Large Hadron Collider. As hadron colliders go, this one is, indeed, large, at just under 17 miles. As far as berth for autonomous vehicles go, on the other hand, its 1.5-inch inner clearance isn't a lot to work with.
Ahead of the Embedded World conference (and NVIDIA’s big show), Arduino is dropping its first major new robotics board since being scooped up by Qualcomm last year. Built around the Dragonwing IQ8 process, the Ventuno Q microcontroller features 16GB of RAM, and 64GB of storage, with a focus on robotics and AI applications. The name, for the record, is Italian for 21 — a nod to the fact that Arduino is now of legal drinking age here in the States.
As we noted in earlier coverage, Qualcomm’s big play in robotics is three-fold. First, the company is leveraging its existing work in autonomous driving and drones to branch out into more physical AI applications. The industrial robotics side of the business has already moved on to Dragonwing IQ9. Second, the purchase of Arduino is aimed at getting in early with developers and startups by powering a popular line of microcontrollers. Third, the company is partnering directly with the companies themselves. Just this week, it announced a deal with Germany’s Neura.
Dexterity was physical AI before physical AI was cool. In fact, it was physical AI before we collectively settled on that term (I always liked “embodied AI” – it reminds me of a nice, warm cup of coffee). Speaking of naming things, the company has slapped the label “Foresight” onto its massive enterprise logistics model. As it notes in a blog post, what sets this system apart from a lot of world models is the fact that it’s been training with real packages in real warehouses in the real world for a total of 100 million autonomous actions.
Mobile manipulation has been a big piece of this physical AI talk of late, and Dexterity puts the issue pretty bluntly, referring to other models as “insufficient for robots that need to touch things.” The company adds, “Manipulation demands a fundamentally different kind of understanding. When a robot picks up a box, the world changes. Objects shift. Previously hidden surfaces become visible. Stability conditions evolve. The world model must not only perceive these changes: it must predict them before they happen, reason about them as they unfold, and update its beliefs based on what actually occurred.”
Swiss firm RIVR has just unveiled the latest version of its wheeled-legged delivery robot. As evidenced by the above photo, RIVR 2 is built for more than just food. It's also designed for last-mile parcel drop-offs. The shot features someone utilizing the bulky system's top-loading mechanism. Once the robot reaches its destination, it sits down, releases the payload onto the ground, and then moves along its merry way. Given that last bit, it's probably easier to avoid thinking about RIVR 2 as a "robot dog."
Would Harold and Kumar have become a true buddy comedy classic, had it starred a pair of hapless delivery robots? What form would a robotic Neil Patrick Harris take? These are just some of the questions I've wasted my morning on, after Uber Eats announced plans to deliver White Castle via Serve Robot. While the burger joint is largely a midwestern and New York metro phenomenon, deliveries will be limited to those places in which White Castle and Serve's existing service already overlap.
Over the years, I've developed a fondness for earnest corporate quotes in fast food press releases. I can't really say why, but here's White Castle Sr. VP of Restaurant Operations, Chris Shaffery, “At White Castle, innovation has been at the heart of our business for more than 100 years. Partnering with Serve Robotics gives us an exciting new way to combine convenience, technology, and great taste together, while allowing Cravers to enjoy their favorite items in a fun and sustainable way.”
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.
.jpg?upscale=true&width=1120&upscale=true&name=R0001754_(1).jpg)
