Best of LinkedIn: Robotics CW 40/ 41

00:00:00: Provided by Thomas Allgaier and Frennus based on the most relevant LinkedIn posts about robotics in calendar weeks, forty-and-forty one.

00:00:07: Frenness is a B to B market research company that supports robotics and smart manufacturing providers with building feature by feature competitive intelligence.

00:00:16: That shows exactly how their product stacks up against the competition.

00:00:19: You can find more info in the description

00:00:22: Right.

00:00:22: so we've got a ton to cover today.

00:00:24: Yeah We really do.

00:00:25: And um To kick things off I want you picture something.

00:00:30: Picture a hundred thousand dollar humanoid robot, right?

00:00:34: Being lowered into a vat of boiling steel.

00:00:36: Oh wow like a movie

00:00:38: exactly Like Arnold Schwarzenegger standard right there in full protective gear watching it melt.

00:00:43: It sounds totally like a sci-fi scene but that actually happened last week.

00:00:49: is this highly calculated corporate decision?

00:00:51: That's wild

00:00:52: Right.

00:00:53: so welcome to the steep dive.

00:00:55: If you're a smart building manufacturing professional, whether on the shop floor or managing operations.

00:01:00: This one is well it's specifically for You.

00:01:04: definitely

00:01:04: our mission today has to cut through all that viral hype of like backflipping robots and look at The real top robotics trends from across LinkedIn.

00:01:13: We are strictly focusing On how physical AI Is moving out Of the lab And onto the actual messy production line.

00:01:20: Yeah because the narrative Has just completely fractured.

00:01:22: At this point I mean, for years the whole industry was just asking like can a humanoid physically do this one complex

00:01:29: task?

00:01:29: Right.

00:01:30: There's demo videos

00:01:31: Exactly but we aren't asking that anymore.

00:01:33: The hardware has crossed a threshold.

00:01:35: Now the question is Can we build ten thousand of these things reliably?

00:01:39: Yeah it really feels Like the model T moment For robotics...I mean the innovation isn't Just the machine itself any more It's assembly line building the Machine

00:01:49: Precisely.

00:01:50: But I mean, are we actually seeing that scale in reality?

00:01:53: Oh it is happening right now.

00:01:55: Thomas Midekis and Nikhil Kumar shared some insights on this about UB Tech.

00:02:00: They basically just finalized the world's first true mass delivery of humanoid robots.

00:02:04: Wait, mass delivery like how many

00:02:07: We're talking?

00:02:08: hundreds Of their Walker S-II units going to automotive And smart factory partners.

00:02:13: The orders crossed eight hundred million yuan.

00:02:16: Wow okay.

00:02:17: so well over a hundred million dollars completely changes the math.

00:02:21: It does, but um... The craziest part is their facility in Liuzhou.

00:02:25: That factory is set up to produce over ten thousand humanoids a year.

00:02:30: Oh

00:02:30: my god.

00:02:30: Yeah

00:02:31: that breaks down two like one new humanoid rolling off the line every Ten minutes and the Walker s-two units are literally on the floor helping build the other robots.

00:02:40: So robots building robots.

00:02:42: Exactly, they handle material movement and their engineer to swap their own batteries so that can run twenty-four seven.

00:02:48: no humans needed to plug them in.

00:02:50: That is just explosive compounding growth when the build empower themselves?

00:02:55: Yeah And the macro data backs it up.

00:02:57: John Coitzie noted that humanoid shipments jumped.

00:02:59: uh four hundred thirty two percent In the first half of twenty twenty six alone.

00:03:04: That's insane!

00:03:05: That

00:03:05: wild.

00:03:06: yeah

00:03:06: who's leading

00:03:07: Mostly?

00:03:08: AGI Bot and Unitry, but Figure actually made the list as first US company.

00:03:13: Nice!

00:03:13: And hardware is just maturing so fast.

00:03:16: Gershon Selnikar compared Boston Dynamics bulky hydraulic atlas from twenty nineteen which you know leaked with super loud to their twenty-twenty six all electric version.

00:03:26: it's totally smooth now.

00:03:28: Twitter also dropped there.

00:03:30: sunny humanoid in Cassie Mobile Robot The physical chassis basically becoming a polished commodity.

00:03:35: Okay, well let me challenge that just a bit.

00:03:38: If the bodies are so polished why is the end of the arm still such a bottleneck?

00:03:43: Because Guillermo Flor brought up at point Elon Musk made That no robot hand can currently drive a simple screw into a piece of wood.

00:03:52: Yeah...that's just The brutal reality of physics.

00:03:54: Driving a screw requires spatial awareness and variable force.

00:03:58: We do it instinctively But for a robot It's total nightmare to replicate Right Especially trying to manufacture ten thousand hands that can do it.

00:04:08: So what's the actual hold up in manufacturing them?

00:04:11: It is all about tactile sensing.

00:04:14: Leo Sue broke down a Tesla patent for scaling flexible tactile sensors To give hand feeling, you need these complex three D structures

00:04:22: Like printing circuits on curved surfaces.

00:04:25: Right

00:04:26: If print them flat and stretch them thermo forming them The circuits break.

00:04:30: but printing directly onto a three-d surface is insanely slow and expensive.

00:04:35: So the hands are still a massive point of failure, which actually brings me back to that melting robot because you know what happens when these things break or get obsolete?

00:04:46: They aren't just steel right

00:04:48: exactly.

00:04:49: Leonard Rodman shared that story about figure melting their retired f dot o two robots.

00:04:54: those were The ones they used at BMW but they had the f.o.four coming out and maintaining old ones was draining resources.

00:05:00: But, They can't just throw them in a dumpster.

00:05:02: No way.

00:05:03: The actuators in those joints hold highly sensitive IP.

00:05:06: so figure ran A melting simulation practiced drops onto airbags In San Jose And found specific foundry that could safely melt lithium ion batteries.

00:05:16: That is brilliant.

00:05:17: It protects the ip AND gets some amazing PR.

00:05:19: Yeah, and we'll see more of this as humanoids expand.

00:05:21: I mean AAA-Twenty launched Bob for Las Vegas corporate events And Dynamic Creatures debuted Hugo which has forty three actuators for theme parks.

00:05:30: Well so they're everywhere.

00:05:32: But okay if the body is getting cheaper in easier to mass produce then The bottleneck definitely shifts to the brain right?

00:05:38: We have talk about physical AI models.

00:05:40: yeah the software layer Is everything now

00:05:42: but i've got a pushback on This.

00:05:44: If these Models are Getting So Smart Processing so much data Are we gonna hit like a thermal or battery wall way before we hit our reasoning wall?

00:05:52: AI compute takes so much power.

00:05:54: You're totally right to ask that.

00:05:56: Garry Tripathi actually highlighted this power dilemma, real-time robot control needs run at about eighty three hertz to keep the robot from falling over.

00:06:04: Eighty Three updates in seconds.

00:06:05: Right but current AI inference The Reasoning part only runs at about zero point two to zero point four hertz.

00:06:12: Oh wow that is a massive gap.

00:06:14: Huge!

00:06:15: Bridging that requires so much compute, in mobile robots right now the GPU draws over thirty seven percent of total power while physical motors only draw sixteen percent.

00:06:25: Wait really?

00:06:25: Yeah

00:06:26: The brain thinking about moving uses double energy at legs.

00:06:30: actually moving.

00:06:31: Yes exactly Which?

00:06:33: why everyone is racing to build Vision Language Action Models, or VLAs.

00:06:37: Microsoft Research just launched a new open-weight five B model called ROWE.

00:06:41: Oh I saw that mentioned by Michael Murray and Andre Kolobov and Ashley Lawrence.

00:06:47: Yeah they pointed out.

00:06:48: the row separates learning robot from learning task.

00:06:52: They do this embodiment mid training

00:06:54: Meaning the robot learns its own body limits first Exactly.

00:06:57: And because they separate it cuts fine tuning data needed for a new task in half.

00:07:03: That's like, um not teaching a teenager complex basketball plays while they're still learning to walk in a straight line.

00:07:08: That's the perfect analogy.

00:07:10: and this separation is making generalization skyrocket.

00:07:13: Z Peng Fu noted that Reward AI's OM-One model is hitting zero shot generalization across totally different arms than humanoids.

00:07:21: Okay Zero Shot meaning it sees new object just infers how to grab without being trained on.

00:07:26: It

00:07:27: learns directly from human data, with zero tele-operation.

00:07:31: You don't need engineers in VR suits puppeting the robot anymore.

00:07:35: DeepMind is doing this too!

00:07:37: Carolina Parada and Rohan Thakur showed Gemini for Argonne teaching Aptronics Apollo Robot to play Cornhole end-to-end.

00:07:45: So robotics has officially become a reasoning problem

00:07:48: Which means we need software tooling.

00:07:51: I know Swami Siva Subramanian and Oscar Tuhumchu talked about AWS launching an open source physical AI tool chain on NVIDIA stack.

00:07:59: Yeah, the infrastructure is arriving.

00:08:01: Shri Ella Prolu found that one in seven startups has already building physical AI

00:08:06: and The way they're building it is wild.

00:08:08: Lars n Joceler brought up in pie.

00:08:11: It uses a I coding agents to autonomously improve the code for robot fleets.

00:08:15: Oh right

00:08:17: like it writes this script runs a physical trial, reads its own failure logs and rewrites the code to fix it's mistakes.

00:08:24: It hit a ninety-nine percent pass rate in just eight attempts.

00:08:27: That is incredible.

00:08:30: But you know even with autonomous coding You still need an ocean of physical data.

00:08:34: You can't just scrape text from the internet for this.

00:08:36: Right, you need real-world physics.

00:08:38: Exactly.

00:08:39: Tim Cardin highlighted a Munich startup that got massive seed funding to pay twenty thousand everyday people To film themselves doing basic physical tasks.

00:08:49: That's their training data.

00:08:50: Oh thats smart.

00:08:51: And sometimes its not even AI right?

00:08:53: Leaf Yang noted how cheap one hundred thirty eight dollar stereo camera beat way more expensive competitor simply because of physics.

00:09:01: It had a wider baseline, so it has four times better depth precision for humanoids.

00:09:05: Yes sometimes the simple hardware hack wins!

00:09:08: But this actually brings us to the most painful part of this whole transition... The deployment reality check?

00:09:15: Yes because having smart brain in shiny chassis doesn't mean anything if its stops assembly line.

00:09:21: Exactly.

00:09:22: Real quick by the way If you're finding this useful You should probably just hit subscribe right now So you don't miss our future additions on how these deployment strategies actually shake out.

00:09:31: On the factory floor,

00:09:32: definitely

00:09:33: but okay let's talk about the ninety-nine percent reliability trap.

00:09:37: I think of this like baseball.

00:09:38: right if a baseball player hits point nine I know meaning they only strikeout every one hundredth pitch.

00:09:46: They are literally a legendary Hall of Famer Right But an assembly line.

00:09:51: that one strikeout halts the entire factory.

00:09:54: Yeah, Nicholas Savage broke down the math on this.

00:09:57: if a robot has a ninety nine percent success rate per action and A job requires one hundred consecutive actions that Robot will only finish The Job thirty six point six percent of the time

00:10:08: which is terrible.

00:10:09: a human would get fired for it.

00:10:10: Thirty-six percent success right

00:10:12: completely And that's why Savage says robotic recovery is everything.

00:10:17: The robot has to sense its own mistakes like dropping a bolt and fix it in real time,

00:10:21: which is so hard to do.

00:10:23: I mean Robert Little highlighted Walmart struggles with automation.

00:10:26: They rebuilt warehouses canceled shelf scanning robots scrapped their pickup towers.

00:10:32: just painful expensive lessons

00:10:34: Yeah But they persist because they have two an Aaron pray through me.

00:10:37: at a great point.

00:10:38: the US doesn't have a robot shortage.

00:10:40: We haven integration and skills gap

00:10:43: right.

00:10:43: So where do these humanoids actually fit in?

00:10:46: Jan Ziska analyzed GXO, the massive logistics company.

00:10:50: They piloted forty-five humanoids and you know how many went into production

00:10:54: out of forty five.

00:10:59: maybe ten

00:11:00: zero

00:11:01: wait zero.

00:11:01: why?

00:11:02: because the real value isn't repetitive.

00:11:04: picking bolted down automation.

00:11:06: does that better?

00:11:07: The true value of a humanoid is exception handling.

00:11:10: Picking up a dropped box, refilling a random supply station... Handling the chaos so other machines keep running.

00:11:17: So it changes whole workflow?

00:11:19: Gabriel Millian pointed out that physical AI shifts.

00:11:21: question from what can we automate to which judgments do you keep with people?

00:11:26: Exactly!

00:11:27: Operations leaders have to redesign roles around exceptions.

00:11:30: Alec Alekbarov actually advises starting with jobs.

00:11:33: people dread like shoveling snow and just giving the robot an existing human tool rather than redesigning the facility.

00:11:40: Yeah, but to use a human shovel The robot needs to know what it feels Like.

00:11:44: that's similar.

00:11:45: real gap is tough

00:11:46: right.

00:11:47: Florian Palatini talked about the URAI trainer Recording actual force and torque data from humans guiding robots so it learns literal feeling of the task.

00:11:56: And Ralph Golden noted how Siri Act uses post-deployment continual learning.

00:12:01: They train the robot on an actual job site to deal with unpredictable lighting and edge cases.

00:12:06: Speed is critical there too.

00:12:07: Rudy Cohen mentioned InVolt, raising eleven million euros... ...to cut industrial robot deployment from months down just minutes!

00:12:15: Which is huge because corporate buyers are still really hesitant.

00:12:19: Lucas M. Ziegler pointed out that corporate investors rarely actually buy their robots.

00:12:24: they invest in.

00:12:25: Schaeffler is a rare exception.

00:12:27: They ordered over a thousand humanoids, but from a company they didn't invest in.

00:12:31: Excellent robotics right?

00:12:33: Okay so if deployment is this hard who's funding all of these infrastructure?

00:12:37: let us talk about the global race because I honestly look at it and wonder is Shunk's eighty-year old family owned approach to building simple robot hands The ultimate proof that boring hardware Is actual moat In AIH?

00:12:51: It s great question But first Look At Capital Divide.

00:12:54: Nicholas Shlopsna shared that Europe actually has eighty-five physical AI companies compared to China's sixty five.

00:13:00: Oh wow, Europe is ahead in volume.

00:13:03: Yeah But they trail massively and capital the average US company as a hundred forty eight million dollars in backing.

00:13:09: in Europe It's just fifty three million.

00:13:11: That's

00:13:12: huge gap.

00:13:13: Natalie Loccianova did note that European funding jumped a hundred and ninety percent in H-one, but it's super concentrated in mega rounds like Nura Robotics.

00:13:22: Still the ecosystem in Germany and Switzerland is solid right?

00:13:25: Erdemirin and Gianmaria Spetta pointed out Germany has six robotics unicorns like Robco and Agile robots despite the automotive slump.

00:13:35: And Marcus Wagner highlighted that Switzerland creates six times more VC-backed robotics startups per capita than Germany.

00:13:41: They just lack the late stage capital,

00:13:44: which brings us back to Schunk.

00:13:46: Sasha Weiler brought them up there this highly profitable boring German family business quietly dominating the hardest part of robotics The hand.

00:13:56: no flashy PR No VC money.

00:13:59: Just relentlessly solving hardware problems.

00:14:01: That's a huge mode

00:14:02: definitely But we talked about dropping boxes in a warehouse.

00:14:08: What happens when the margin of error is human life?

00:14:11: Healthcare,

00:14:12: yeah The stakes are completely different.

00:14:14: and Tom Holland mentioned strikers.

00:14:16: four hundred million dollar investment into the UK's CMR surgical.

00:14:20: He was worried.

00:14:21: UK innovation just gets sold to the US before scaling

00:14:24: right And while the surgical robot market is projected to hit twenty-seven billion dollars by twenty thirty, according to Mayuri Lamba.

00:14:31: Sebastian Kaliri pointed out that whole sector only raised one point one billion dollars last year...

00:14:36: Which was basically one funding round for a humanoid company?

00:14:39: Exactly!

00:14:40: But there are lower risk applications.

00:14:42: David Rieger and Andre Dachchenko launched Nura HealthTech.

00:14:45: using robots just move empty hospital beds.

00:14:49: It returns over thirteen hundred care hours to nurses facing labor shortages.

00:14:53: That's such perfect use case.

00:14:55: It's low risk, but high value.

00:14:58: Unlike surgery where strict safety limits are terrifyingly critical Dr.

00:15:02: Michael Menegin analyzed FDA data and found one hundred forty four deaths tied to robotic surgery.

00:15:08: Wow He used that viral video of a robot flawlessly kicking or remote as an example.

00:15:14: if the robot does exactly what it is told But at the wrong millisecond in an operating room its catastrophic.

00:15:20: Yeah, Pascal Bourne had a dark funny take on this.

00:15:23: Surgery needs robotic precision but you absolutely do not want your robotic surgeon freezing at thirty-seven percent to ask if you accept cookies mid incision.

00:15:32: Exactly!

00:15:33: It has be governed by human judgment

00:15:34: For sure.

00:15:36: Well as we wrap up I wanna leave with the paradigm shifting idea from RJ Pittman and Misha Dohler.

00:15:41: They argue that As physical AI scales Real estate itself will transform into configurable nodes from machine operations.

00:15:51: The physical AI cloud?

00:15:53: Yeah, and consequently the telecom industry will actually stop charging for bits delivered... ...and start charging for intelligence in actions delivered at the edge.

00:16:03: That is going to completely rewrite the Telecom business model.

00:16:07: Absolutely Well.

00:16:09: if you enjoyed this episode new episodes drop every two weeks.

00:16:13: Also check out our other editions on smart manufacturing and industrial AI digital construction and connected tools in equipment.

00:16:21: Thanks for listening, don't forget to subscribe!