I’m going to keep the intro short this week. It’s been a difficult one for personal reasons. For now, I’m going to encourage folks to donate to the Michael J. Fox Foundation, if you have the desire and means to do so.
Likely my trip to Automate is going to be a bit shorter than originally planned, but I hope to see some of you in Chicago next week, particularly during the Humanoid Robot Forum and — of course — Thursday’s live Automated recording with FieldAI’s Ali Agha.
EU folks, I will be hosting a stage at the upcoming Davos Tech Summit, on July 2 and 3. And then I’m in Paris for a day or two, covering the Machina Summit.
As for the lead image in this newsletter, Automated is currently in the throes of a giveaway. Subscribe to our YouTube channel (you’re likely already subscribed to the newsletter if you’re reading this) and fill out this form to enter to win robot lego kits from Serve, Cobot, Dexory, and Matic, along with some slick Automated stickers and an enamel pin that will make you the envy of tens of people.
Fun stuff in this week’s issue, including a look back at Anki with Bedford Robotics CEO Boris Sofman and a chat with Genesis AI’s VP of commercial and strategy, Vivian Sun, about the physical AI firm’s new humanoid-ish robot. Also, definitely don’t miss this week’s podcast with Aya Durbin, as we discuss all of the hard, unsexy parts of scaling robots in an industrial setting. 😎
There’s an alternate version of Boris Sofman’s career that plays out differently. In this timeline, the CMU robotics PhD gets more heavily involved in the day’s DARPA challenge and — like many of his classmates — finds himself immersed in the early days of self-driving startups immediately out of school.
Most timelines lead there eventually. In ours, Sofman would ultimately spend just shy of five years at Waymo, the last three and a half of which he served as the Alphabet subsidiary’s senior director of engineering, head of trucking. In March 2024, he and a handful of others exited the company to form Bedrock Robotics, applying many of their learnings to the world of heavy construction machinery.
What makes Sofman’s career fascinating, however, is the road it took along the way. There was an internship at iRobot before the internship at Neato Robotics that first brought him out to the San Francisco Bay Area. The latter also evolved into a four-year role at the robotic vacuum firm, leveraging the sorts of autonomy and path-planning concepts that formed the foundation of his thesis.
In early 2010, Sofman and fellow CMU Robotics PhD candidates took a crack at a field of consumer robotics, a field that was (and remains) even more treacherous than self-driving in certain respects.
“We had a pretty good thesis in that we were seeing smartphones come out and start to really scale,” Sofman tells me. “Android and iPhone started ramping into the hundreds of millions of units, which drove down the cost of every single component that they had. So microcontrollers, memory, cameras, motors, wireless chips. These are all the building blocks of consumer products. We had the realization that, if you're smart, these prices are dropping 10x off of what they normally were, and you can start to do things that were not cost-competitive previously. On top of that, you can start to use smartphones as the brain and interface to the physical world. We had that insight like really early on.”
July 1 marks a year since Genesis AI emerged from stealth with $105 million in seed funding. Examining the size and volume of massive physical AI funding rounds that have emerged in the intervening months, it’s easy to lose sight of how massive that amount is — particularly when dealing with the seed stage.
It’s true that the unfathomable sums of money being funneled into the space are the product of substantial hype surrounding concepts of generalized AI, but there’s another phenomenon worth examining here. A major motivating factor is a bid by many to launch a full-stack solution.
In my consumer electronics days, we would frame the conversation in terms of iOS versus Android. It’s massively expensive upfront to build a stack, but the theory goes that building hardware and software under the same roof creates a more unified experience. Your own mileage will vary, but Microsoft and Google have clearly seen enough value in the concept to attempt similar models with Surface and Pixel, respectively.
Robotics has historically been more fragmented for myriad reasons, but — particularly in a world of general-purpose systems/humanoids — firms are increasingly seeing the value of the full-stack approach. This marks a shift in philosophy from a robotics world that was far more likely to embrace off-the-shelf components. But again, that’s a robotics world where funding rounds in the hundreds of millions of dollars were unheard of.
It should be no surprise that many of these well-funded physical AI firms are taking their own crack at general-purpose systems. Rhoda AI is a good example of a top-tier firm that has announced its intentions to launch a humanoid in the near future. Today, Genesis AI is doing precisely that with Eno, a “next-generation robot” the company says is “the culmination of Genesis AI’s full-stack approach to building robots that operate seamlessly across industrial and consumer environments, with human-level capabilities.”
Humanoid robots are everywhere in the headlines. But Aya Durbin says the real test is not whether a robot can impress people in a demo. It is whether that robot can deliver real value, positive ROI, and reliable performance inside industrial environments.
Andrew Barry (Generalist AI) - Generalist AI's cofounder and CTO, Andrew Barry, discusses physical AI's difficulties and jaw-dropping breakthroughs.
Daniel Rausch (Amazon) - What does Alexa look like in the age of large language models? Amazon VP Daniel Rausch discusses how the smart assistant is continually evolving.
Daniela Rus (MIT CSAIL)- This episode has it all: physical AI, sperm whale alphabets, and some cutting edge insight into autonomy. It's also recorded in front of a live audience at MassRobotics. You're not going to want to miss it.
Agents are beginning to play a role in generating the code required to engineer robots based on data feedback. As this post by NVIDIA notes, however, they’ve largely been in the realm of digital environments, whereas real-world applications require more direct human input. The team is showcasing a new system called ENPIRE, which is designed to create a feedback loop to continually develop and refine these systems’ efficacy. The name is acronymish: “Environment module (EN) for automatic reset and verification, a Policy Improvement module (PI) that launches policy refinement, a Rollout module (R) to evaluate policies with single or multiple physical robots operating in parallel, and an Evolution module (E) in which coding agents analyze logs, consult literature, and improve training infrastructure and algorithm code to address failure modes.”
ENPIRE is designed to run on fully autonomous robots. Those behind the project, which also include researchers from CMU and U.C. Berkeley claim the system already has a “99% success rate on challenging, dexterous manipulation tasks in the real world, such as PushT, organizing pins into a pin box, and using a cutter to cut a zip tie.”
China’s Cloud giant, Alibaba, has expanded the reach of its Qwen software offering into the world of physical AI. The newly announced Qwen Robot Suite centers around a trio of offerings. There’s the video world model Qwen RobotWorld, the vision-language navigation model Qwen RobotNav, and Qwen RobotManip, a VLA designed for real-world manipulation, which is trained on 38,000 hours of open-source training data. The South China Morning Post (another Alibaba property) notes that the move is a build to expand the region’s reach into physical AI, after thus far largely highlighting the hardware aspects of humanoid robots.
We — and by “we,” I mean “an oversubscribed event with roboticists, engineers, designers, students, kids, grandparents” (my invite must have gotten lost in the mail) — got our first Weave’s new robot, Isaac 1, on Friday. As you can no doubt surmise from a quick glance, Isaac 1 is the softer, gentler, more commercially viable version of the $8,000 home robot we saw earlier this year. The system shares the same rough form factor as the Isaac 0, with a more polished design that amps up the human connection and lowers the potential pinch points. This social media post from the company shows the robot doing a fine job picking up and putting away stuffed sloths, which may or may not be an issue you’re presently grappling with in your own home.
If you really think about it, the vast majority of startup stories begin with founders not going to work on a Christmas tree farm. This one is just a bit more explicit on that front. After all, PerPlant’s CEO, Rasmus Emil Hansen, is the son of a Danish Christmas tree farmer. And while he didn’t follow in his father’s footsteps exactly, Hansen is applying his time to farming. Founded in 2022, the Copenhagen-based agriculture ag-tech startup builds AI cameras that can help target spraying when applied to heavy machinery. “I wanted to go back to my roots and do something that is actually more meaningful and more important, impact-wise, something I care truly about,” Hansen told Becca in this piece. “How can we help farmers prepare better food in a more sustainable way and use new technologies to do it?”
Teaching a system to navigate an unfamiliar space is hard enough when it has a full view of its surroundings. But what’s a poor robot to do when it’s expected to perform a task in a new environment of which it only has a partial view? A joint production by researchers from Johns Hopkins, Lambda, and the University of Cambridge introduces the notion of “3D belief,” which in the system creates a conception of what the world might look like behind some of those corners. “At each step,” the researchers write, “the agent updates its belief from egocentric RGB history and camera poses, samples multiple plausible 3D beliefs, imagines observations along candidate paths, and selects the rollout with the strongest open-vocabulary goal progress under uncertainty.”
Johns Hopkins researcher Yifan Yin notes on social media, “We argue that an agent needs more than view synthesis or video prediction. It needs an explicit belief over the 3D world: what it has observed, what may exist beyond view, and how that belief should update as it moves.”
Off the top of my head, I’m not sure of the practical applications of a system that allows you to train a robot to unsheathe a sword while you’re blindfolded. I’m not saying that they don’t exist — they just don’t immediately spring to mind. Like many other robot demos in the social media age, however, such scenarios seem to be more about proving a point than any practical application. In the Universal Manipulation Exoskeleton (UME) Stanford and Alibaba-owned Ant Group, that specific application is a way to showcase the value of real-time haptic torque feedback when training the upper bits of a robot via tele-op.
A bit more practical in real-world terms are things like picking drinks out of the fridge and even flipping boxes. The system can also be applied to robots that look even less human than the substantially abstracted model above, including more standard arms. The above bi-dexterous mobile manipulator was built for just shy of $10k (specifically $9,533), using the OpenArm 1.0 system, actuators from Damiao Technology, the Hexfellow PCW-25 power casters for the base, and 2K HDR cameras from Jieruiweitong. Did I mention you can teach it to unsheathe a sword while blindfolded? Unnecessary, sure, but you never know.
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=Aya%20Durbin%20(2).png)
.jpeg?upscale=true&width=1120&upscale=true&name=images%20(3).jpeg)
.jpg?upscale=true&width=1120&upscale=true&name=1f8f4582-4f68-4cd7-90a4-871ddbe30309%20(1).jpg)
