Best of LinkedIn: Robotics CW 42/ 43

00:00:00: Provided by Thomas Elgair and Frenis based on the most relevant LinkedIn posts about robotics in calendar weeks, forty two and forty three.

00:00:08: Frenes is a B to be market research company that supports robotics and smart manufacturing providers with building feature-by-feature competitive intelligence That shows exactly how their product stacks up against the competition.

00:00:21: you can find more info In the description.

00:00:24: so right now um in an auto parts plant in Hongzhou there's this robot working a twenty-four seven shift, handling something like sixty thousand parts across fifteen different stations.

00:00:35: And it never takes the coffee break?

00:00:37: Exactly!

00:00:37: It doesn't have.

00:00:38: you know...shift handovers and every single time that moves its quietly vacuuming up data to make the robots standing next to it just little bit smarter

00:00:47: which is wild thing

00:00:48: about right?

00:00:49: So if you are a smart build-in manufacturing professional, this is the deep dive for you today.

00:00:54: We are tracking the top robotics trends surfacing across LinkedIn and well The biggest takeaway is that physical AI has aggressively moved out of the laboratory in onto the actual factory floor.

00:01:04: Oh absolutely we're looking at massive capital deployment real world pilots And quite frankly a very intense regional turf war.

00:01:12: Yeah, the shift is just.

00:01:13: it's undeniable

00:01:14: It's palpable.

00:01:14: yeah

00:01:15: I mean four years.

00:01:16: the industry was basically captivated by conceptual models or, you know carefully edited videos of robots doing backflips.

00:01:22: The Myral Parkour videos

00:01:24: exactly.

00:01:26: but what we are seeing in this recent data is a total pivot toward commercial commitments.

00:01:31: the engineering talent and more importantly the capital Are heavily concentrated on?

00:01:36: What is actually shipping and generating measurable value today?

00:01:40: Well, I want to start right there with the capital because the sheer volume of money sloshing around the ecosystem is staggering.

00:01:47: It really does!

00:01:48: We are talking about global venture funding into robotics hitting sixteen point three billion dollars and that's just for the first quarter of twenty-twenty six alone.

00:01:58: Yeah...to put that sixty point three million in perspective it actually surpasses total funding the robotic sector saw from all of twenty twenty one to twenty twenty four combined.

00:02:08: Wow.

00:02:09: And a massive portion of that is flowing directly into humanoid development.

00:02:14: Startups building humanoids pulled in a record eight point six billion dollars across one hundred and four different funding rounds just this year,

00:02:21: That's incredible!

00:02:22: And Neuro Robotics has leading the pack

00:02:30: at one-point-four

00:02:37: billion.

00:02:38: But I was looking at the RobCo data and that stood out to me for a totally different reason.

00:02:42: Oh, with those secondary sale?

00:02:44: Yes they just hit a one billion dollar valuation but they did it through a secondary sale worth around forty million dollars

00:02:50: right which is a massive signal of market maturity.

00:02:53: how so

00:02:54: well.

00:02:54: earlier this year robco raised hundred million dollars in a primary round two.

00:02:58: you know build-the-business.

00:02:59: A secondary sale means early employees.

00:03:01: investors are actually selling their shares To new investors

00:03:05: so they're cashing out a bit.

00:03:06: Exactly, They are turning paper wealth into actual liquidity.

00:03:09: When employees can buy houses with their robotics equity it proves the financial ecosystem around this hardware is highly functional and demand is sustaining.

00:03:17: Yeah, that makes sense.

00:03:18: And we're also seeing physical scale play out.

00:03:20: in Emeryville California Foundry Robotics just closed an over-subscribed one hundred million dollar series A and they are backing up with a one hundred thirty thousand square foot facility to build an autonomous factory.

00:03:32: So it's not software running on the server It is pouring concrete moving actual metal.

00:03:37: We can't ignore underlying model data layers pulling huge rounds to support that metal either.

00:03:43: Definitely

00:03:44: You have dream labs exiting stealth with ninety million in seed funding.

00:03:48: Echo Robotics pulled in one hundred and fifty millions specifically for a vision force action model, right?

00:03:54: And Mecha AI raised to sixty million series B With the very specific focus just recording human motion data from body sensors To teach robots how we move

00:04:04: Right because a vision force action model isn't just about the robot seeing a bolt, right?

00:04:08: It's about the Robot feeling the physical resistance as it turns that Bolt and you know adjusting its Force in real time.

00:04:15: Exactly.

00:04:15: But what really caught my attention is how strategic equity is shifting into real commercial commitments.

00:04:21: like Scheffler placing a binding order for one thousand to two thousand humanoids by twenty thirty-two.

00:04:27: That's a huge bet.

00:04:28: And Mercedes actually put Eptronic units through a grueling twelve-month pilot on their floor before they committed to investing.

00:04:35: They aren't just writing checks anymore, they are kicking the tires incredibly hard!

00:04:39: Well manufacturers are demanding proof of value now... I mean in a decade ago A flashy demo might secure a partnership?

00:04:46: Right

00:04:47: Today, plant managers require the hardware to integrate into existing workflows demonstrate safety around human workers and you know prove an eventual return on investment before any binding orders are signed.

00:04:58: I hear that.

00:04:59: but then i look at physical intelligence.

00:05:01: oh yeah

00:05:02: they are currently in talks to raise at least one billion dollars at a twenty billion dollar valuation which is double their prior eleven point five billion dollar evaluation from earlier this year.

00:05:12: all right And here's The Kicker.

00:05:13: They Are Doing This With Absolutely Zero Revenue.

00:05:17: Yeah, how do we reconcile that?

00:05:19: We just said manufacturers demand proof of value.

00:05:21: But here is a company with zero revenue commanding a twenty billion dollar price tag.

00:05:25: It feels exactly like the dot-com boom but for physical atoms are these valuations completely disconnected from reality?

00:05:33: it isn't totally valid skepticism and on The surface it looks absurd

00:05:39: right.

00:05:39: But the reason a pre-revenue company can command a twenty billion dollar valuation is because the broader ecosystem is proving that hardware is finally scaling.

00:05:48: Okay.

00:05:49: Investors are looking at factory deployments happening across industry and basically pricing in the inevitability of technology.

00:05:56: Physical intelligence isn't selling robots, they're building foundational brain

00:06:01: like operating system?

00:06:03: Exactly!

00:06:03: The OS for physical movement.

00:06:05: The market is betting that whoever builds the default Windows OS for humanoids will capture trillions in value, which makes a twenty billion dollar bet today seem well rational to venture capital.

00:06:17: So the money is flowing into the brain because the physical metal was already moving out of the door?

00:06:21: Yes.

00:06:22: And when we look at who is moving that medal, the dominance is just striking.

00:06:27: In the first half of twenty-twenty six Chinese manufacturers accounted For over ninety seven percent Of global humanoid shipments.

00:06:34: The scale of their manufacturing base gives them an incredible advantage.

00:06:37: I mean, they're output of humanoids tripled year over year.

00:06:41: but the truly critical metric here isn't just production volume it's deployment.

00:06:46: Over seventy percent.

00:06:47: those units are actually deployed in real businesses doing actual work rather than sitting at university research labs.

00:06:55: It is.

00:06:56: Analysts projecting five hundred thousand active units by twenty thirty.

00:07:00: based on this trajectory

00:07:02: I want to focus on where these robots are going, because automakers are essentially turning into robot makers.

00:07:07: Yeah they really are.

00:07:08: Back in twenty-twenty one only two carmakers were seriously building humanoids.

00:07:12: Today nearly twenty of them Are deeply involved.

00:07:16: XPENG's robotics division just raised over nine hundred million dollars at a six point three billion valuation.

00:07:22: Toyota co led a three hundred million dollar seed round for Walden Robotics and They've had walden robots working On their own factory floors since February.

00:07:30: twenty twenty six And Mitsubishi Motors is exploring humanoid production at their Kyoto plant for twenty-twenty seven.

00:07:36: well automakers have realized that a car factory Is the perfect dojo?

00:07:40: For a humanoid robot.

00:07:42: How, do you mean?

00:07:43: they already possessed The global supply chains the precision engineering talent and the complex multi stage assembly environments required to stress test these machines.

00:07:52: I see

00:07:53: by testing robots on Their own lines automakers get A dual benefit They optimize their own throughput and they incubate a massive new hardware product that can eventually sell to other industries.

00:08:03: Let's

00:08:03: dig into what the throughput actually looks like, because I was looking at recent deployment data for the Leju Quavo wheeled humanoids

00:08:09: The ones in Hangzhou.

00:08:11: Yes!

00:08:11: The plant we mentioned earlier.

00:08:13: It has handled over sixty thousand parts across fifteen stations and logged about a thousand hours.

00:08:18: The data shows a ninety-eight percent loading success rate

00:08:21: Maintaining a ninety eight percent success rate.

00:08:24: live dynamic manufacturing environment.

00:08:26: It's a massive engineering hurdle.

00:08:28: to clear it proves the reliability is really there,

00:08:31: it does.

00:08:31: but let's look at this cycle time because The cycle takes the robot about forty seconds To complete right?

00:08:37: A human worker on that exact same line Does the cycle in about fifteen seconds.

00:08:42: now two a plant manager listening?

00:08:43: Forty seconds probably sounds totally unviable.

00:08:46: Yeah too slow

00:08:47: exactly if the human is nearly three times faster How does this deployment actually make any financial sense on the factory floor.

00:08:54: This comes down to rewriting the traditional math round OEE, overall equipment effectiveness.

00:08:59: we naturally bias toward burst

00:09:01: speed.

00:09:02: right

00:09:02: a human does the task in fifteen seconds.

00:09:04: but human workers are biological systems.

00:09:06: they take coffee breaks eat lunch they stop to talk experience fatigue towards end of shift and you know eventually go home

00:09:14: Right!

00:09:14: They have to sleep.

00:09:15: Exactly That forty second robot cycle is continuous.

00:09:19: It runs a twenty-four seven shift without single microstoffage.

00:09:23: Over an hour, the human wins!

00:09:26: But over a week?

00:09:28: Over a month or year... The slower infinitely consistent robot completely dominates total throughput calculation.

00:09:35: Consistency beats burst speed in manufacturing.

00:09:38: Its the tortoise and hare but applied to industrial automation

00:09:42: Precisely

00:09:43: That mechanism makes total sense.

00:09:44: And we are seeing that consistency went out in other pilots too, like Agility Robotics is expanding its digit deployment to sites in North Carolina and Tennessee after moving one hundred thousand totes.

00:09:55: yeah an AGCO just signed a commercial agreement to bring Digit into their Kansas factory in twenty-twenty seven.

00:10:01: well as these fleets expand it creates a massive ripple effect for component suppliers.

00:10:05: two oh

00:10:06: sure

00:10:07: building humanoid isn't just software.

00:10:09: requires incredibly complex hardware to mimic human joints actuators.

00:10:14: The motors and gearboxes that act as the robot's muscles currently make up fifty to sixty percent of the entire cost of a humanoid,

00:10:20: And put specialized suppliers in highly lucrative position.

00:10:24: companies like Regal Rexnard Timkin and SKF are perfectly positioned for this.

00:10:28: definitely regal rex Nord actually saw their discrete automation orders grow by more than nine percent recently driven heavily by distraction and humanoids.

00:10:37: but I also find HD Hyundai robotics approach fascinating

00:10:40: though with the wheeled robot.

00:10:41: yeah Instead of fighting the incredibly difficult physics of bipedal walking right away, they are sequencing manipulation before locomotion.

00:10:49: Right They built a dual-arm manufacturing robot that rolls on wheels.

00:10:53: It can carry seven kilograms per arm has highly dexterous five finger hands and they're essentially saying let's perfect The complex hand movements first And we can figure out balancing in two legs later.

00:11:04: it is a highly pragmatic approach.

00:11:06: You know while humanoids capture the imagination Conventional robotics are still quietly breaking records.

00:11:12: Right, they aren't going anywhere!

00:11:13: Exactly the International Federation of Robotics reports there are now five million industrial robots operating globally an increase of nine percent and China's dominating that installation base as well accounting for fifty-nine percent all new industrial units.

00:11:29: But there is a physical ceiling to how fast hardware can move.

00:11:32: I mean, you can optimize actuators all day but to get that forty second cycle time down to fifteen seconds the robot has stop hesitating.

00:11:41: It needs think faster.

00:11:42: Exactly it need adapt variables instantly and brings us into the software bottleneck where real battle for brain happening.

00:11:52: By the way, if you want to stay on top of how these physical AI models are evolving week-to-week make sure You hit subscribe so you catch our future deep dives.

00:11:59: Yeah because The evolution On the software side is moving at a breakneck pace.

00:12:03: Microsoft research recently released row which Is an open weight five billion parameter vision language action model built on their Phi family right?

00:12:12: Their architectural approach is what makes it stand out.

00:12:15: They explicitly separated learning the robot's specific physical mechanics from learning the general logic of the task.

00:12:22: Let's break that down because I think its crucial, Its like a difference between learning how to operate a specific car pedals and steering wheel versus learning actual rules on road.

00:12:31: Thats great analogy

00:12:32: By separating the two Microsofts model as outperforming competitors on major benchmarks Because It doesn't have to relearn How To Think every time its put into new robot body

00:12:41: Exactly the right framing and neuro robotics is pushing a similar boundary with neuron AI.

00:12:48: It's a perception language action model that integrates three-d vision touch audio And language,

00:12:54: it works across different platforms.

00:12:56: Right?

00:12:56: Yeah, they were to cross their humanoids robotic arms in mobile manipulators.

00:13:01: They also partnered with nxai to integrate what they call episodic memory.

00:13:06: Episodic memory on a factory floor is huge.

00:13:08: It means if the robot strips a bolt or struggles with a glare on a part, it remembers that specific episode Wednesday and approaches the task from different angles.

00:13:22: And we are seeing general-purpose language models being tested directly as these robot policies too.

00:13:28: Yeah, the data shows that GPT-Six Astra reached a sixty six point three percent success rate on the Robolab one twenty benchmark when queried directly.

00:13:36: Which is pretty impressive!

00:13:37: The progression there was fascinating... When they applied an energetic setup with GPT Six Astra That's accessory jumped significantly from sixty six points three percent up to seventy four point eight percent.

00:13:48: Okay, wait.

00:13:48: I have to stop you there because when i saw that in the data The cost implication blew my mind.

00:13:53: Oh!

00:13:53: The cost per episode?

00:13:54: Yes

00:13:55: When we say an agentic setup We mean...the AI isn't just answering a prompt.

00:13:59: It's acting as an autonomous agent.

00:14:01: it looks at the goal writes its own step-by-step plan Executes it checks Its own work and corrects its mistakes In real time.

00:14:08: Right..it is constantly querying itself And That

00:14:10: level of constant self querying takes massive compute.

00:14:15: Doing that raised the costs Per Episode from a dollar ninety to an eye-watering, thirty six dollars and forty cents.

00:14:22: It's a massive jump.

00:14:23: Thirty six bucks for a single action?

00:14:25: I have to call that out.

00:14:27: there is no plant manager on earth who will pay thirty six Dollars For A Robot To Pick Up A Single Part.

00:14:33: Is This Just An Expensive Laboratory Parlor Trick Right Now?

00:14:37: Well At Thirty Six Dollars In Action Yeah it Is Completely Unviable For Any Real World Factory Production Line Today.

00:14:43: You Are Entirely Right To Push Back On That.

00:14:45: Okay Good

00:14:46: But What we have to understand is that the developers aren't paying thirty six dollars just to get the task done.

00:14:53: So what are they paying for?

00:14:53: They're

00:14:54: paying thirty-six dollars to generate high fidelity data of how the robot solved a complex problem autonomously.

00:15:01: That thirty six dollar action becomes perfectly labeled highly complex training example, it has fed back into smaller cheaper models To make them smarter.

00:15:09: Ah

00:15:10: I see

00:15:11: It reveals that true bottleneck in physical AI right now Is not GPUs...it's the data

00:15:16: High quality data of physical movement in the real world.

00:15:20: Exactly That is what Physical Intelligence was doing when they pre-trained their Pi Zero model on roughly ten thousand hours of proprietary robot data.

00:15:28: Right The real breakthrough, Is that the deployed fleets themselves are becoming massive data vacuums.

00:15:34: Sariax Cortex Model is running over two hundred live systems across Europe

00:15:39: And generating Real World Data.

00:15:40: Exactly Because those systems are alive They have accumulated one billion real world picks.

00:15:46: Every single time a human operator has to intervene and correct the robot, that fallback becomes a labeled correction.

00:15:52: The deployed hardware basically acts as a distributed sensory network.

00:15:56: The more robots you have failing in being corrected into real world faster your central model improves

00:16:01: Which creates some massive moat.

00:16:03: Yeah And this data loop explains why we are seeing such radical shift In underlying infrastructure of industry.

00:16:09: That's why NVIDIA acquired hugging face for twelve point nine billion dollars.

00:16:14: Acquisition is wild when you think about the implications.

00:16:17: NVIDIA isn't just selling pickaxes for AI gold rush anymore, they are buying land, mines and railways.

00:16:23: They're vertically integrating the entire path.

00:16:26: Right!

00:16:27: They provide GPU silicon in server farm.

00:16:29: now own hugging face model hub with its eighteen million developers all down to physical compute on walking robot.

00:16:37: Amazon makes aggressive horizontal moves as well.

00:16:40: AWS launched open source physical AI tool chain built on top of Nvidia stack.

00:16:46: Which makes sense,

00:16:47: but Amazon brings a unique asset to the table.

00:16:50: They operate more than one million robots in their own fulfillment centers.

00:16:54: Wow, right?

00:16:54: Right they

00:16:55: are using that unparalleled operational experience to build a tool chain designed to let other manufacturers launch advanced capabilities In weeks rather then spending years building custom infrastructure.

00:17:06: and the telecom industry is jumping into the infrastructure play too.

00:17:10: Oh

00:17:10: T-Mobile.

00:17:10: Yeah t-mobile is utilizing five g advance to solve a major hardware problem edge computing.

00:17:17: Think of it like a chef in a busy kitchen, okay?

00:17:19: The robot's onboard chip acts as the chef's reflexes catching a falling knife instantly.

00:17:24: that requires millisecond local control so the robot doesn't drop a heavy car part if the Wi-Fi blinks

00:17:30: right.

00:17:30: safety is local

00:17:31: Exactly.

00:17:32: But the five G connection is the chef's massive cookbook located up in The Cloud, where all of heavy power-hungry reasoning and planning happens.

00:17:41: T-Mobile moving that higher level reasoning to network edge which means robot doesn't have carry a fifty pound supercomputer on its back.

00:17:49: just think

00:17:50: It is brilliant division between local safety and cloud intelligence.

00:17:54: And you know this infrastructure build out highlights how different regions are approaching robotics ecosystem.

00:18:00: Yeah, the strategies are very different.

00:18:02: Very?

00:18:03: The US is dominating in hyperscale capital and cloud infrastructure.

00:18:07: China is dominating raw manufacturing volume and supply chain clustering.

00:18:11: right

00:18:11: but Europe has playing a very different highly strategic game.

00:18:15: Europe is heavily investing in the physical testing infrastructure Right like Nura and the Technical University of Munich recently opened RoboGym.

00:18:23: yeah robo gym is fascinating.

00:18:25: It's a massive space with around one hundred robots operating across forty-eight training cells.

00:18:30: it allows industrial companies to actually bring their proprietary production processes into the gym and test them on hardware before buying.

00:18:38: That

00:18:38: is huge value

00:18:39: add!

00:18:40: And has backing of heavy hitters like Bosch, Scheffler & NVIDIA.

00:18:44: Well The Industrial Foundation in Europe particularly Germany is incredibly deep

00:18:48: for sure.

00:18:49: Germany has rich legacy.

00:18:50: precision engineering currently hosts deep bench robot arm makers.

00:18:55: Companies like KUKA, EGUS, RobCo, Fruitcore, Agile and Nura.

00:19:00: They have what?

00:19:00: six robotics unicorns right now?

00:19:02: Yeah

00:19:03: Six Robotics Unicorns.

00:19:04: What stands out in the data is the web of deep industrial partnerships they are forming to deploy this technology

00:19:11: Like The Bosch Partnership.

00:19:12: Exactly We saw seven major partnerships noted recently including Bosch Integrating with Nura Renault Group Partnering With Hexagon Robotics And Audi Working With Agile Robots.

00:19:23: And that precision engineering translates to European design being highly influential globally.

00:19:29: The microduck robot, which was designed in Bordeaux by teams from Polyn Robotics and Huggingface actually became the most popular robot of twenty-twenty six selling over fifteen thousand units.

00:19:41: It's an impressive feat!

00:19:42: The stakes globally are incredibly high right now.

00:19:45: I mean, Anthropics S-One filing explicitly named robotics and advanced manufacturing as major growth markets for their foundational models.

00:19:54: But this brings up a really critical question about vowel capture for Europe.

00:19:58: Zurich is a massive incubator for robotics talent.

00:20:01: you have ETH Zurich Disney Research the RIA Institute.

00:20:05: they are producing some of the most brilliant minds in.

00:20:08: Yet we see companies like Atterion moving their headquarters to the US.

00:20:12: We see Robcoz, CEO relocating to the

00:20:14: U.S.,

00:20:15: to follow the market.

00:20:16: It's a trend for sure.

00:20:17: Why is Europe so successful at building brilliant minds and physical infrastructure but struggles to keep capital on the market value from bleeding across The Atlantic?

00:20:26: It is classic structural dilemma of European tech!

00:20:30: The European ecosystem excels in deep-tech research long term precision engineering establishing thoughtful regulatory frameworks.

00:20:38: But the US ecosystem excels at rapid commercialization, high risk tolerance and access to hyperscale capital.

00:20:44: The hyperscale capital is key?

00:20:45: Yes

00:20:47: When a robotics company clears the research phase And suddenly needs a hundred million dollar check To scale a factory or deploy massive fleet U.S venture capital market Is simply structured to write that cheque faster.

00:20:58: So the talent is born there, but money made here.

00:21:01: Exactly!

00:21:02: The intellectual talent is incubated in Europe... ...but financial value ultimately captured by the US.

00:21:08: It's a fascinating dynamic to watch especially if you are tracking where true power centers of this industry will settle over next decade.

00:21:17: We have covered the explosion of capital and reality of humanoid outperforming humans on continuous throughput.

00:21:23: The software brains tackling data bottleneck and fierce global turf war.

00:21:28: But you know, as we look at all these breakthroughs there is one final massively unresolved bottleneck buried in the data that I think every smart manufacturing professional needs to be aware of.

00:21:39: Okay what?

00:21:39: Is it

00:21:40: well We have discussed robots executing highly complex reasoning utilizing episodic memory and following plain language instructions.

00:21:48: right.

00:21:48: The brains are getting huge

00:21:50: but despite All That algorithmic brilliance the most physically capable systems In the world still Have To Stop Every Few Hours To Recharge

00:21:58: The battery problem.

00:21:59: Hardware constraint meets software ambition.

00:22:02: precisely

00:22:03: charge autonomy is the silent killer of OEE.

00:22:07: AI reasoning, especially agentic setups requires massive compute power and Massive Compute drains batteries incredibly fast

00:22:16: because they're basically carrying a server rack.

00:22:18: Yeah almost no.

00:22:19: current systems on the market score high on both advanced reasoning and Charge Autonomy.

00:22:25: If you were looking for the next massive under-invested opportunity in smart manufacturing, it is solving the power bottleneck for physical AI.

00:22:34: That makes total sense

00:22:35: because a genius robot with perfect episodic memory isn't generating any value if its spends half of shift plugged into the wall.

00:22:43: What a phenomenal point to end on!

00:22:45: A Genius Robot Isn't Useful If It's Plugged Into The Wall.

00:22:48: Exactly

00:22:49: If you enjoyed this episode, new episodes drop every two weeks.

00:22:52: Also check out our other editions on smart manufacturing and industrial AI digital construction and connected tools in equipment.

00:23:00: Thank you so much for joining us on the steep dive And remember to hit subscribe So you never miss an addition.