An artist recently told me she “[loves] having a lot of chapters in [her] life.” True to form, she’s an artist, cartoonist, musician, mom (her stuff's here, if you'd like to see it). She’s the kind of person who makes for a great interview (for my other podcast), due to a breadth of knowledge and general restlessness that always sets in sooner than later with creative types.
I’ve turned this framing over in my head a few times since we spoke. While depth of varied experience is something I’ve long valued in others, I think I’ve written it off when reflecting on my own work.
I’ve had a somewhat unconventional career — here we are, for instance, discussing my music podcast in my robotics newsletter. For better and worse, it hasn’t remotely resembled the one I assumed I was embarking upon as a creative writing major graduating from a liberal arts college in a bear market. The truth is no one really has the career they expect — though some certainly come closer than others.
When I reflect back on my own career, it can be hard for me to discern the narrative structure. Life is strange, markets are volatile, tastes are always changing, and technologies can still surprise. Sometimes the jobs we imagine for ourselves growing up no longer meaningfully exist by the time we get there. If we’re lucky, however, something even more engaging has cropped up in its place.
Ali Agha’s is the kind of career I think of when I think about chapters. DARPA challenges, space helicopters, cave exploration, the search for extraterrestrial life — it’s a hard book to put down. Makes for an engaging podcast, as well, and I’m happy to finally be releasing our live conversation from Automate in June as this week’s episode.
I’m sure I don’t have to sell anyone reading this on a career in robotics, but these sorts of conversations are a reminder of the many fascinating directions this space can take you, if you’re up for the challenge. It’s a bit like the piece below about the CMU team that took their snake robots from a stage with the “Semi-Charmed Life” guys to the aftermath of Venezuelan earthquakes in the space of 36 hours. Expect the unexpected. Doo, doo, doo, doo, doo-doo, doo.
Lots of unexpected this week, like a metric ton of startups you’ve never heard of — including one from Uber’s cofounder. We’ve got a quick catch-up with the Generalist guys about their latest model and a conversation with Sudo AI’s COO we recorded outside a Swiss supermarket.
Sunday Robotics spent its nascent weeks amid a sea of 3D printers stuffed inside a rented house in Mountain View. A stone’s throw from Google and the Stanford campus where its cofounders both studied, the small team had set up shop in the heart of Silicon Valley.
That doesn’t mean they were ready to admit to the neighbors they were building robots. In a launch video, founding designer Shy Yang says the startup told anyone who asked that they were running “an Etsy business.”
The real story of Sunday’s founding is a tale of two research papers from Stanford roboticists. While working on researching complimentary projects at the same school, it took an encounter on X to make Tony Zhou aware of Chen Chi’s work.
The former, who now serves as Sunday’s CEO, worked on ALOHA alongside Stanford teaching staff, Sergey Levine and Chelsea Finn (both of whom are now members of Physical Intelligence’s large founding team). The project focused on the development of a low-cost teleoperated mobile system capable of bi-dexterous manipulation.
Chi, meanwhile, had recently showcased UMI (Universal Manipulation Interface), a project from Stanford and Columbia that also featured future Walden Robotics founder Russ Tedrake. The work utilized electronic “gloves” designed to capture physical data for robot training. It’s a modality that gained increasing traction among startups over the past year.
As both of these complimentary projects focused on home scenarios, it was, perhaps, inevitable that their fake Etsy store/robotics startup would take a similarly unstructured path. I’ll not belabor the numerous reasons such a use case is considered a non-starter for much of the industry, but when we recorded an episode of the Automated podcast last week, Zhou pushed back on many of the orthodoxies framing home use as a far-off bet.
If you had to make a grocery run in Davos, Switzerland on July 3 or 4, there’s a fairly decent chance you were greeted at the door by the above robot. For two days, Sudo AI set up camp in the front of a Migros supermarket location, kitty-corner from the Congress Centre, as it served as the focal point for the inaugural Davos Tech Summit.
The station was part of a small-scale robot takeover spanning a two-block radius amid a cluster of luxury hotels. Sudo’s was among the more interactive exhibitions, positioned as a “challenge” for event attendees and unsuspecting locals alike. Shoppers could pick out any item (within certain parameters, one imagines) and witness R1’s display of dexterous manipulation.
Following our conversation on a bench outside the store, cofounder and COO Martina Hansen was quick to point out that — despite the display — Sudo is not a robot company. Not exactly. “Rather than building another robot company,” she says, “we’re building the developer platform that enables many robotics companies to build faster.”
More specifically, the company is focused on simulation, data, and infrastructure for building, testing, and deploying third-party systems. For now, however, the company does, in fact, build robots. I know this for a fact, as I watched one drop an apple in a bin mere minutes before our conversation.
Sudo is currently a full-stack operation, which is to say it’s not “another robot company” that happens to build robots. This is owed, in part, to the company’s Shanghai location, which puts it in close proximity to people, parts, and machines that manufacture and assemble such systems.
“We find from a lot of lab researchers they say they need good robot hardware to build on,” says Hansen. “Sometimes they're not even interested in our foundation model. They want to work on that themselves, but they still need a good platform to build with, and there are not that many great solutions out there right now. So, we think we can serve that market need as well.”
Agha’s path to FieldAI runs through MIT, Qualcomm, NASA JPL, and DARPA. At Qualcomm, he worked to put autonomy on the low-power Snapdragon processor that would later fly aboard NASA’s Mars helicopter. Watch on YouTube>
Samantha Johnson (Tatum Robotics)- Johnson's ambitions to connect the deafblind grew as the community was cutoff from contact during the pandemic.
Ayanna Howard - The incoming president of Spelman College has based her illustrious career around creating robotics for good, from pioneering climate work at NASA to a STEM startup.
Yoel Fink (MIT)- Materials engineer and longtime MIT professor, Yoel Fink, discusses what it means to live a life based on first principles.
The Federal Communications Commission Tuesday announced the addition of “advanced robotic devices” to its “Covered List.” The document is a section of the Secure Networks Act designed to ban systems that “pose unacceptable risk to the national security of the United States or the security and safety of United States persons.” Additions in 2021 banned the import of Huawei and ZTE devices, causing substantial disruption to the smartphone manufacturers’ global businesses. The July 28 addition cites, “Foreign-produced advanced robotic devices, except advanced robotic devices which have been granted a Conditional Approval by DoW [Department of War].” You can find more on the exemption process here.
Reporting from Reuters notes that the bans “target imports of new Chinese robots and power inverters, seeking to protect the U.S. AI buildout from national security threats and reshore key industries slated for explosive growth.” These bans are expected to disproportionately impact companies like China’s Unitree, which has disrupted the industry in recent years with its low-cost robotic systems. Other nations will be affected by this decision as well. Humanoids and quadrupeds are notably singled out in the press release tied to the announcement. The net is expected to be considerably larger, however, including household items like robot vacuums.
Uber co-founder Travis Kalanick is betting big on robotics and physical AI with Atoms, a company he's been building for ~8 years, per a16z's Ben Horowitz. The venture capitalist had some modest words about the firm, noting, "Some companies are inevitable. The moment we invented the computer, it became inevitable that computers would take over the entire world of bits. Computers inevitably did the job of transforming information, sending information, and storing information, for every application in the world, in order to save us time and make us more productive."
Atoms has thus far raised $1.7 billion in equity investment funding led by Andreessen Horowitz. In his own post, Kalanick says the germ of the company actually dates back 16 years, to Uber's inception. More to the point, he is taking a more seemingly more traditionalist approach in his thinking, championing the specialized robot — a tried and tested model that has fallen a bit out of fashion with Silicon Valley, as VCs rush to fund generalized physical AI and humanoid form factors to match.
"Builders need to be careful of the human fascination with seeing mechanized versions of us," Kalanick says, before peppering a bit of Moravec's Seasoning™ in. "Every humanoid demo or launch in recent memory usually starts with a humanoid dance show or martial arts display. The recent humanoid Olympics in Beijing highlighted many advances in humanoid development. I watched the half-marathon and couldn’t help but think how much better it would be if they just had wheels. Gainfully employed robots are the machines best suited for the job at hand that can make a living doing it."
Following the recent news of an OpenAI model hacking Hugging Face, NVIDIAhas joined a broad coalition of companies in forming the Open Secure AI Alliance. Hugging Face is among those taking part, while OpenAI's name is conspicuously absent from both the list and the announcement post. Other launch participants include: Adobe, Cadence, Capital One, Cisco, Cloudera, Cloudflare, Cognition, CrowdStrike, Databricks, Dell Technologies, DoorDash, Elastic, HPE, Hugging Face, IBM, LangChain, the Linux Foundation, Microsoft, NAVER, NetApp, Nous Research, OpenClaw, Palantir, Palo Alto Networks, Red Hat, Reflection AI, Salesforce, SAP, ServiceNow, Siemens, SK Telecom, Snowflake, SpacexAI, Synopsys, Thinking Machines Lab, and TrendAI.
A post announcing the news warns against "blanket restrictions" from regulators and lawmakers. "As policymakers and regulators grapple with AI safety, it will be crucial to recognize open models, harnesses and security tooling as defensive assets, not liabilities, in AI and cybersecurity policy," the company writes. "Blanket restrictions on open frontier AI systems would weaken defensive capacity and risk concentrating power, dependence and vulnerability in a few closed providers. Companies and governments should invest in shared open infrastructure for AI defense — datasets, evaluation frameworks, attack simulators and red-teaming tools — much as past generations invested in open source software."
If there’s one thing uniting the entirety of the physical AI industry at present, it’s an insatiable quench for that sweet, sweet data. The conversation tends to focus on factories, warehouses, and homes, as these form a kind of holy trinity for general-purpose robotics at the moment. But the truth of the matter is everything we want robots to do with flexibility at scale will require massive troves of data, and certainly the $12 trillion construction industry is no exception.
Or, as Gritt told Aaron for his recent piece on the newly announced startup, "There is no existing dataset for construction. You can't recreate a 10,000-acre jobsite in a laboratory. The only way to build that intelligence is to deploy on real projects and do real work." Gritt has raised a total of $32.4 million in funding, the lion’s share of which comes from a recent $26 million Series A. The Silicon Valley-based firm is building a software layer to bring autonomy to existing heavy construction equipment.
Mimic entered the physical AI scene with a bang, announcing a high-profile Audi deal not long after emerging out of stealth. Becca explores another key partnership for the ETH Zurich spinout — albeit one that ultimately grabbed fewer headlines, owing to the lack of participation from a leading German luxury automaker. Freiburg-based Black Forest Labs brings a lot to the table, as the pair team up to release a new vision language action (VLA) model, FLUX-mimic. "You could argue that [Black Forest] fundamentally invented modern AI image generation,” Mimic cofounder Elvis Nava tells A3. “I applied a lot of that research into robotics and that's kind of how we got started in inventing the whole application of this aspect to robotics. When we [got] into contact with them again this year, it became very obvious that it was a very nice match.”
"Within approximately 36 hours, our team was on stage performing with Third Eye Blind, finished, and then assembled the robots, cameras, protective equipment, tools, and replacement components and traveled to the affected region,” CMU professor Howie Choset told Liam during an interview for this piece. “We went because the technology was potentially useful, but also because we believed it was important to show up and serve when people were facing an extraordinary tragedy."
There's certainly some cultural whiplash, going from showcasing your robots on stage at an American semiquincentennial birthday event alongside a 90s Buzz Bin band to the rubble of an earthquake-ravaged coastal city 30 minutes north of Caracas. If anything, though, it goes to show how many worlds these technologies touch, shifting from on-stage entertainment to search and rescue a continent away over roughly a day and a half.
You’ve heard of sim-to-real — likely in the context of the associated gap that occurs when models trained on purely synthetic data are deployed in the real world. How about real-to-sim? This one’s pretty much what it sounds like, referring to real-world data/settings used to simulate synthetic twins for a digital testing environment. Okay, now what about real-to-sim-to-real (R2S2R)? I suppose this one, too, is more or less the way it sounds, albeit with a level or two of extra complexity, as synthetic scenarios developed with real-world components are, in turn, used to train real-world systems.
SceniX is focused on this multi-hyphenate work it is now doing as part of the Fei-Fei Li-led World Labs, as part of an acquisition announced last week. This week, the pair are showcasing the fruits of their recent collaborations, noting in a new blog post, “Using our proprietary technology, our model generates simulations aligned with reality that allow robots to do what has long been considered out of reach: learn complex manipulation tasks with zero real-world training data; predict and evaluate through simulation which robotic policies will succeed or fail in the real world, without extensive and expensive physical trial and error; and then operate reliably for hours on physical robots in real-world settings.”
There’s some consensus that generative AI’s recent evolution is set to profoundly impact the way robots learn. There is, on the other hand, plenty of debate around the specifics. Among the most frequent topics is the precise role large language models (LLMs) will play in all of this. Proponents have long pointed to natural language commands as a strong contender. Y Combinator-backed Waddle Labs proposes taking things a step further, developing what it has deemed the “Claude Code for robotics.”
The team breaks it down thusly: “By using agents to control robots, we combine the generalization of VLAs with the speed and reliability of classical approaches. Our agents decompose a goal into subtasks, and completes each subtask by viewing camera feeds, writing control code, and calling models like VLAs. The agent outputs a program that you can run and iterate on by talking to our agents.” Waddle claims its technology can create a working policy like "Write a program to package the box" in as little as 20 minutes.
“Company buys one device” generally doesn’t pique my interest enough to warrant inclusion in the roundup. Were that the case, this already very long edition would almost certainly never make it to your inbox. I do, however, have a special interest in Virtuix, having followed the company since their earliest days, when I demoed an early prototype in a Los Angeles hotel room more years ago than I care to mention. “Jury-rigged” is putting it nicely, but the idea was there, and the company has seen its vision through to reality with its passive VR treadmill. There’s probably a book to be written about Virtuix’s journey to this point, though — having left the world of consumer electronics — I haven’t really followed the last few chapters.
Suddenly, however, the startup has crossed my radar yet again by selling its first “Enterprise system” to Tesla’s Optimus Robotics division. Says Virtuix, “The purchase highlights Virtuix’s continued expansion into enterprise and humanoid robot applications, following a series of previously announced milestones across consumer, defense, and healthcare markets, including collaborations with Meta, NASA, and all four major branches of the U.S. military.” Certainly it presents a compelling modality as humanoid companies look to VR for pre-training, post-training, and system fallback.
The team behind Scythe Robotics just launched Agency Tool Company. The Y Combinator startup is building software for robot fleet deployment with partners including Burro, Tempo Works, and Gather AI.
Tacta is building robot hands powered by "fluid tendons." Tacta itself is powered by a $75 raise, led by Toyota's Woven Capital.
Enigma raises $71 million to "make using robots as intuitive as using a laptop or smartphone." Meantime, you can use a robot arm to paint on an easel via your browser.
Humanoid design is getting weird, and I'm 100% here for it. Deux is machina.
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.
%20-%20Edited.jpg?upscale=true&width=1120&upscale=true&name=maxresdefault%20(5)%20-%20Edited.jpg)
.jpg?upscale=true&width=1120&upscale=true&name=mimic%20hands%20800%20(1).jpg)
