Yesterday we dropped the 50th episode of the Automated podcast, which — as my more mathematically inclined colleagues inform me — puts us roughly half the way to 100. Thanks to everyone who has gotten us this far. And to those who have stood in our way, know that vengeance will rain down upon you. Just kidding. We’re cool. (Or are we?)
Our 50th episode is a special one for me. Last week’s big feature on Parkinson’s tech was built around this episode. Kathryn Zealand will tell you that I’ve been following her work closely since her Google X days.
In the intervening years, my dad received his official Parkinson’s diagnosis, and Skip started trials involving the condition. My dad’s condition came up every time Zealand and I have touched base, up to and including this episode. Like Skip, my immediate family is based in the Bay Area, and I always harbored some faint hope that the technology would sufficiently mature to a point where it could directly improve his quality of life.
He passed between the time the interview was recorded and when it posted. While he and I discussed the technology a number of times, his condition was too advanced for him to ultimately benefit from it. Writing about his struggles on forums such as this spurred a number of you to reach out and share how the condition has impacted your own lives.
It’s been made plain to me through these conversations that — much like dementia — Parkinson’s is a condition that impacts nearly everyone. It drives home how important this work is.
I asked Zealand to introduce me to someone on the research side of the clinical trials. You’ll find that conversation with University of California San Francisco assistant professor Jessica Bath on the back half of episode 50. I hope you all get as much out of the conversations as I did.
You’ll find two more recent conversations I appreciated below. The first is my long overdue writeup of my Humanoid Robot Forum panel, "Scaling, Delivering, and Deploying: Humanoids in the Real World.” That chat features Noble Machines’ Elizabeth Samara-Rubio, TRI’s Erin McColl, McKinsey’s Ani Kelkar, and Silicon Foundry/Dexterity’s Rebecca Yeung.
Immediately below that is a feature that spawned from my chat with University of Zurich associate professor Nadine Zurkinden, after her presentation at the Davos Tech Summit. The piece explores how driver liability shifts as cars become increasingly autonomous — and how much catching up regulators still have to do.
Confession time: I got so wrapped up talking about Digit V5’s knees that it kind of overshadowed other tidbits during last month’s interview with Agility cofounder Jonathan Hurst. It’s also worth highlighting another underreported element of humanoid pilots. Per the roboticist, “If the robot fails, we then have to make sure that somebody is still moving that product and making that happen.”
Most of the time that means swapping a malfunctioning unit out with a working one. Pretty standard operating procedure. If push really came to shove, Hurst told me, it could potentially mean that the human in charge of supervising the robot would have to step in and make sure the job got done.
That could mean an engineer suddenly being tasked with moving all of the requisite totes from point A to point B during a shift. Hurst was quick to add — in Agility’s case, at least — that was purely a hypothetical, as early pilot problems could generally be addressed through hardware maintenance.
Two weeks prior to that conversation, I moderated a pair of panel discussions at A3’s Humanoid Robot Forum. We recently went into some depth about our conversation around physical AI and mobile manipulation. The other chat, "Scaling, Delivering, and Deploying: Humanoids in the Real World," covered some more under-discussed aspects of humanoids relevant to this conversation.
“A real positive we're seeing with this specific form factor is that you can just move it out of the way,” Noble Machines’ chief business officer, Elizabeth Samara-Rubio, noted during that chat.
The reason why this generally goes unremarked upon should be obvious. When rattling off a list of your new technology’s chief advantages, the ability to move it out of the way probably isn’t making your top 10, unless you’re selling lawn chairs.
But there is, in fact, an inherent plus to this ability to swap systems in and out. There are larger trends like seasonality and changing tastes, which require supply chain flexibility. But even more practically, there are those instances when a system just isn’t working as intended. In these cases, you may need a human being to tag back in.
“Most automation we're used to requires big trials, proving success, some sign-off process, and then you bolt it to the ground in the factory, and it's expected to run at 99.9%-plus uptime, depending on the application,” Samara-Rubio added. “But because these platforms are small, mobile, and movable, you can run them in shadow mode, and if one goes down, the person who was already doing that job can just step back in. We're seeing the attitude toward deployment shift, because it's genuinely easier to slot these systems into a process — and pull them back out again.”
Our conversation keeps circling back to a slide Nadine Zurkinden highlighted during her Davos Tech Summit presentation. SAE (Society of Automotive Engineers) International’s Levels of Driving Automation chart. First published in 2014 and revised in 2021, the graphic breaks driving autonomy down in five levels — six, if you count Level 0.
Most cars currently on the road fit into the latter category, with features including automatic emergency braking, blind spot warning, and lane departure. Level 5 represents what we’d deem “full autonomy.” There’s a fundamental shift between Levels 2 and 3, between a human actively “driving” to “not driving,” when the driver supported/automated features are engaged.
The chart’s 2021 revision largely clarified the distinction between Levels 3 and 4. Waymo’s fleet of ride-hailing vehicles currently operates at Level 4, without the presence of a fallback human in the driver seat. Crossing into Level 5 means going from features that “can drive the vehicle under limited conditions” to those that “can drive the vehicle under all conditions."
The chart has, predictably, stirred up plenty of debate in the dozen or so years since it was first published, but the standard has provided a rough framework and vocabulary for discussing this technological paradigm shift. No surprise, then, that it popped up again in a July 2026 talk by Zurkinden.
“When I was reading all the legal papers, I was always thinking the acceptable risk part is somehow missing,” the University of Zurich associate professor told me in a conversation following the presentation. “What is the acceptable risk and what are the criminal liability risks that come with driving automation? It's a complex question.”
The SAE graph doesn’t seek to determine criminal or civil liability. At what point does responsibility for failure shift? Broadly speaking, the answer appears to move increasingly in the direction of the company as autonomy increases and human control decreases (though manufacturers no doubt have their own opinions on the matter).
The ability to move is about far more than getting from one place to another. It can shape independence, confidence, social connection, and quality of life. Watch on YouTube>
Andrei Danescu - We visit Dexory's new Nashville testing facilities to chat about the global and vertical expansion of the U.K. startup's 60-foot-tall AMR.
Ali Agha- Live at Automate 2026, FieldAI's CEO discusses his NASA background and the state of physical AI.
Samantha Johnson (Tatum Robotics)- Johnson's ambitions to connect the deafblind grew as the community was cutoff from contact during the pandemic.
Unitree Robotics’ long-awaited IPO prices the humanoid-maker at 150.8 yuan ($22.34) per share, for a total valuation of 61 billion yuan ($9.04 billion), it confirmed Thursday. The valuation is considerably higher than the initial 50 billion yuan/$7 billion valuation reported last fall. The company is aiming to raise 6.1 billion yuan via the Shanghai Stock Exchange listing. AI giant DeepSeek, also based in Hangzhou, is among the chief backers, investing 140.8 million yuan ($20.8 million) per a Shanghai Stock Exchange filing in exchange for 933,399 Unitree shares — or just over representing 2% of the company.
The deal finds both parties entering into a physical AI/hardware partnership. Unitree's near ubiquity as a low-cost offering has found the firm entering into a number of high-profile partnerships in recent years. In June, the company announced H2 Plus, an NVIDIA GR00T reference platform robot designed specifically for academic research. Unitree, which offers a broad range of low-cost robot form factors, has seen massive growth amid the humanoid boom, reporting $235 million in revenue in 2025. Going public will provide additional transparency into those successes, as investors continue to flock to physical AI and robotics investments. In June, U.S.-based Agility Roboticsannounced its own plans to go public. The Pacific Northwestern firm behind the industrial humanoid, Digit, is set to make its stock exchange debut by way of a SPAC merger with Churchill Capital Corp XI.
If you’ve ever wanted to watch a person blindfold a robot while it cuts vegetables, these Dyna videos may be a once-in-a-lifetime opportunity. On Monday, the Bay Area-based company physically took the wraps off its new world model, Dyna-2, which was pre-trained on more than one million hours of first-person egocentric video. Dyna claims the World-Action Model (WAM) is robust enough to transfer learnings across a variety of different robot embodiments, including industrial arms, humanoids, and hands. Per the press material, the startup says it’s able to hit 80–90%task success rates (up from 20%), without requiring additional post-training. “DYNA-2 enables zero-shot performance at production-level speeds and quality across new deployment sites,” the company says, adding that the model “enables physical intuition and spatial reasoning gained from human motion to transfer directly to robot hardware, even without seeing a single frame of robot data in pre-training. This dramatically lowers the barrier to training robots on new tasks.”
I read a bit of coverage about a startup the other day, noting that robotic surgery might be a lot closer than you (the reader) think. I’d say that’s a fair assessment, given that trials for the technology date back to the mid-80s. The da Vinci system received FDA approval for assisted surgery a dozen years later and full surgery by the turn of the century. It’s been 26 years since the system began performing laparoscopic procedures in U.S. hospitals (27 since it hit Europe). These are incredible facts for all sorts of reasons, not the least of which is the fact that — unless you’re the sort of person who subscribes to newsletters like this — there’s a good chance you think we’re still talking about Star Trek stuff here.
I’m happy to report an anecdotal uptick in robot surgery startups. You’ll find one (Andromeda Surgical) in Becca’s funding roundup above. This week she also spoke to the founders of Inner Logic following a $11.5 million raise. Rather than focusing on the hardware itself, the startup has built a platform for iterating and testing surgical and healthcare systems. This concept of procedural and surgical autonomy is clearly arriving and is very top of mind for everybody in the ecosystem,” says co-founder/CEO, Tito Porras. “But everybody's in different stages of adoption from a product perspective. The good thing about what we do is figuring out where the customer is with respect to their adoption of something like autonomy, because autonomy — as it happened in self-driving vehicles — is something that will happen in stages.”
Industrial robot purchases continued to grow, per new data from our parent organization. Unit order increased 4.3% YOY in Q2, working out to a 21.3% increase in revenue. The headline here is an overall diversification in the kinds of systems ordered. A sizable 25% decline in automotive OEM purchases was more than offset by increases to electronics manufacturing, life sciences/pharm/biomed, automotive components, and food/consumer goods.
Massachusetts Institute of Technology debuted research aimed at eliminating the bottleneck present when robotic systems execute vision-language-action models (VLAs). While good at planning out action, data requirements can result in slow and jerky robot movement. The team is presenting a system called VLASH that predicts the future state of the robot and its environment before it has finished the action, thus removing lag time. The process is further sped up by dividing the action into fewer, larger steps. “Even though there is a very large model working in the background," says the paper's co-lead, Jiaming Tang, "VLASH lets the robot react and execute its actions very fast, much more like a human would. This could help to make robots for all sorts of dynamic tasks more effective."
I like to imagine the team behind STEMbot started with the pun and worked their way backward into this plant-climbing robot. Whatever the order of operations, I would like to take this moment to declare my undying love for STEMbot. Long may you live and high may you climb, detecting potential plant-destroying pests as you cruise along. Godspeed, brave STEMbot.
A team of researchers at U.C. San Diego is developing a soft robot designed to navigate brain vasculature, removing blood clots as part of an anti-stroke procedure known as a thrombectomy. The team was awarded $22.9 million from Health and Human Services' Advanced Research Project Agency for Health in order to design a system that could open access to the traditionally cost-prohibitive procedure.
“Our big goal is to make it possible for people in underserved and rural areas to have access to a procedure that we know is very successful at treating strokes,” says associate professor Tania Morimoto. “We will combine embedded mechanical intelligence with artificial intelligence in this soft robot with the goal of making it easier to deploy in essentially any medical setting nationwide, hopefully democratizing access to this time-sensitive procedure.”
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=e031edf5-1218-4ac5-9cd1-714fba12c17a%20(1).jpg)
.png?upscale=true&width=400&upscale=true&name=images%20(5).png)
