Best of LinkedIn: Robotics CW 36/ 37

00:00:00: Provided by Thomas Allgaier and Frenis, based on the most relevant LinkedIn posts about robotics in calendar weeks thirty six and thirty seven.

00:00:07: Frens 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.

00:00:19: you can find more info in description.

00:00:21: So imagine spending billions of dollars to replace your factory workforce with autonomous robots Only to discover you now need to hire thirty percent more human staff.

00:00:31: just keep machines running.

00:00:33: I mean, it sounds backwards but yeah

00:00:35: right that is exactly what has happening at the bleeding edge of logistics Right Now for this deep dive we're looking At The Absolute Top Robotics Trends Across LinkedIn and You Know There's This Massive Glaring Contradiction Staring Us Right In The Face.

00:00:49: Yeah It'S Really The Defining Tension Of The Whole Industry Right Now Because If You Look At The Capital Markets Wall Street Is Valuing humanoid robotics like it's this completely solved infrastructure category.

00:01:01: Like

00:01:01: a done deal?

00:01:02: Exactly, a done-deal.

00:01:04: but then you know... You walk onto an actual factory floor and talk to the engineers And they're openly admitting that most of this hardware still can't actually do reliable paid work.

00:01:14: Right, so you've got venture capital writing these like massive blank checks while the actual physics of a robot just trying to pick up a cardboard box without dropping it are still brutally unforgiving.

00:01:26: Oh

00:01:26: absolutely brutal!

00:01:27: So we really need to cut through the hype today.

00:01:30: We're gonna break down the capital flowing in.

00:01:32: The real production number is actually hitting factory floors that technical bottlenecks they're fighting and this massive sector split in healthcare that most people are just completely overlooking.

00:01:45: Yeah, and the implications for anyone building or managing a smart facility right now?

00:01:49: Are immediate.

00:01:50: you really can't afford to misread where this technology actually sits in its development cycle For

00:01:55: sure.

00:01:55: So let's start with the money because the financial numbers are just Completely detached from the hardware capabilities at this point.

00:02:00: totally detached yeah.

00:02:01: across the market Right Now You've got twelve humanoid startups holding it combined seventy three billion dollars in private valuations.

00:02:10: The U.S.

00:02:11: leads that table, obviously with figures sitting at a thirty nine billion dollar valuation alone

00:02:16: which is just

00:02:18: staggering right.

00:02:19: but China actually holds five of those twelve spots and Europe they only have two Which are nearer a robotics at roughly seven billion?

00:02:27: And humanoid.

00:02:28: But you know You really have to look at how the public markets or validating those private numbers too.

00:02:33: Take the unitary IPO in Shanghai.

00:02:36: yeah That IPO was over subscribed more than eight thousand times.

00:02:40: It closed its first day of trading at a fifty point.

00:02:42: seven billion dollar market cap.

00:02:44: Wait, I have to call it time out here.

00:02:46: fifty billion dollars

00:02:47: Yeah in one day they became the most valuable humanoid maker in the world.

00:02:51: leapfrogging figure

00:02:53: mean for a company making machines that In a lot of cases still stumble around like their walking on ice.

00:02:58: That just feels like pure vaporware To me.

00:03:00: well

00:03:00: it sounds absurd until you realize investors aren't paying For The robot that exists today right?

00:03:06: They're paying to own the infrastructure of tomorrow.

00:03:08: Okay, what do you mean by that?

00:03:10: Look at the compute pact between niscale and figure.

00:03:12: They just committed three point five billion dollars in compute power and that's scaling past six billion to give figure access up to a hundred thousand NVIDIA GPUs.

00:03:23: Oh,

00:03:23: wow!

00:03:23: Okay I see what you're saying.

00:03:24: it's like they are building a million miles of magnetic high-speed rail tracks before anyone has even successfully invented a train that won't derail at ten miles an hour.

00:03:32: Exactly Their stockpiling AI brains for robot body still trips over its own feet Right

00:03:39: And the smartest capital is playing both sides in this timeline.

00:03:42: Take SoftBank For example.

00:03:44: In the exact same window They agreed to acquire ABB or BOTX for roughly five point four billion dollars.

00:03:52: And ABB is an established legacy player, right?

00:03:55: Like they actually make money!

00:03:56: Yeah decades of revenue as a completely proven industrial franchise doing heavy lifting and car manufacturing.

00:04:02: but simultaneously Softbank is looking at a six-billion dollar valuation For majority stake in one X

00:04:08: which the unproven home humanoid bet

00:04:12: Exactly, so they're buying the present with ABB while placing this massive expensive call option on The Future With One

00:04:19: X. But you know building that future exposes an entirely different bottleneck because you can't just construct a physical robot out of venture capital right?

00:04:28: You actually need parts

00:04:30: right.

00:04:30: hardware is hard

00:04:31: yeah and defines investments.

00:04:32: just launched the first us listed robotics actuator etf.

00:04:37: it basically tracks twenty five companies making the physical muscles in joints of these machines.

00:04:42: so like the motors reducers ball screws

00:04:45: legs and shovels.

00:04:46: exactly and mckinsey estimates Those parts alone make up forty to sixty percent of a humanoid as total bill of materials.

00:04:53: So the people selling those picks and shovels are going to make a fortune?

00:04:56: Well,

00:04:56: yeah because hardware is incredibly complex an expensive-to-manufacture at scale which really demands we look what all this capital's actually producing.

00:05:05: We have to separate the declared production goals from reality fleet operations on ground.

00:05:10: Let us talk scale then Because press releases that were just wild.

00:05:15: UB Tech brought their Liuzhu SuperSmart factory online and they're backed by eighteen million in new contracts.

00:05:22: Okay, their target is over ten thousand humanoids a year.

00:05:27: that has literally one robot rolling off the line every ten

00:05:29: minutes.

00:05:30: That's a massive number right?

00:05:32: And meanwhile X-Peng's Guangzhou Line just had its first iron humanoid walkoff under it own power.

00:05:38: They are boasting eighty percent core process automation and targeting mass production By late twenty twenty six.

00:05:44: I mean, those are huge figures.

00:05:46: But if you're actually managing a logistics facility You have to look at the gold standard for automated scale which is Amazon.

00:05:52: Sure

00:05:53: They're running over one million robots across more than three hundred fulfillment facilities via their deep fleet routing system.

00:05:59: Right

00:05:59: they're kings of this...They

00:06:00: are.

00:06:01: but amazon openly admits that there next generation centers require thirty percent more staff in reliability maintenance and engineering roles.

00:06:08: See i am still hung up on that stat.

00:06:11: You fully automate your facility and your maintenance headcount goes up by a third.

00:06:15: Yeah, because physical machines break down they require constant calibration parts replacement And just general troubleshooting is not set in forget.

00:06:22: Then I really have to push back on these startup claims Because if Amazon is struggling with maintenance on like wheeled Roomba's?

00:06:30: essentially how on earth as a company?

00:06:32: Like Agibot realistically claiming their own pace to produce fifteen thousand complex walking humanoids this year.

00:06:40: That's

00:06:41: a great question!

00:06:42: Like, what are they even doing with fifteen thousand unproven units?

00:06:45: Well...they aren't putting them into work.

00:06:47: They're building real-world data factories.

00:06:49: Data

00:06:50: factory?!

00:06:50: Yeah An analyst at the World Robot Conference in Beijing just laid it out perfectly.

00:06:55: They estimated that fifty to seventy percent of humanoid built in China will not do single hour paid productive work.

00:07:04: Wait really?

00:07:05: Seriously Instead, they're just going to sit in these dedicated facilities generating training data.

00:07:09: Wait so the robots can walk around bump into things and fail?

00:07:14: Yes because you realize that you cannot rely entirely on synthetic training from computer simulations.

00:07:21: I mean...in a simulation A box is a perfect cue,

00:07:25: right?

00:07:25: But in the real world.

00:07:27: The boxes dented that tape is peeling and the center of gravity shifts when you pick it up.

00:07:32: so You need them messy physical friction data of the real-world to train the next generation models.

00:07:38: That is wild.

00:07:39: They're literally harvesting failure, but you know to be fair there are practical milestones hitting the floor right now.

00:07:45: Agility Robotics just revealed digit five after racking up over sixty-five thousand fleet hours across nine customer sites

00:07:52: which Is real verified work

00:07:54: exactly it's Now got a nine minute fast charge and an OSHA compliant fifty pound lift capacity.

00:07:59: And then Universal Robots launched its AI-ready Gen Seven platform at IMTS Chicago.

00:08:05: Oh,

00:08:05: and don't forget Skilled!

00:08:06: Yes,

00:08:06: Skilled this one is staggering.

00:08:08: they crossed a hundred million dollars in annual recurring revenue just ten months after their first commercial deployment.

00:08:14: I've already got over sixty paying customers.

00:08:17: I mean skills.

00:08:17: revenue timeline Just proves the immense demand.

00:08:20: like if you can make the hardware actually function The market will throw money at ya.

00:08:25: Absolutely Speaking of making things function.

00:08:28: Quick side note for you listening, if you are finding this breakdown useful make sure to hit subscribe so that we don't miss our future deep dives.

00:08:36: We love pulling these insights out of the noise but getting back into the factory floor... If the software is getting smart why do data collecting humanoids still drop boxes?

00:08:47: Because we're slamming in to technical frontiers of contact physics!

00:08:51: Moving a digital arm on screen is just math But gripping oddly shaped object that's slipping from your fingers in real time, That is an incredibly difficult physics problem.

00:09:04: And it really comes down to latency right?

00:09:06: There was this fascinating post by Entepatrex about this.

00:09:09: Oh yeah the inference step breakdown Yeah

00:09:11: he argued a standard hundred millisecond inference step which you know sounds basically instantaneous to human actually leaves robot completely blind for ninety eight milliseconds of real physics.

00:09:22: Exactly!

00:09:23: Because when an AI model takes a tenth of second to think about its next move, gravity doesn't hit pause.

00:09:31: So if the heavy part starts slipping, ninety-eight milliseconds of blindness means it hits the floor and you break your component or worse – you injure human workers.

00:09:40: How do we even fix that processing gap?

00:09:43: Well, Patrick suggests the Fix isn't using these massive computationally heavy AI context windows for every single micro movement.

00:09:51: The fix is deterministic reflex loops running at five hundred Hertz.

00:09:55: So it's kind of the difference between a brain and a spinal cord?

00:09:58: Exactly,

00:09:58: like if you touch a hot stove You don't wait for your brain to process the temperature in calculator response Your spinal cord just yanks your hand back instantly.

00:10:06: Yes

00:10:06: that has perfect analogy.

00:10:08: these robots need a localized spinal cord for their joints so they Don't have to ask the main AI brain for permission To Just you know tighten A grip.

00:10:16: That makes so much sense.

00:10:17: And It's not just the processing latency either It's the physical sensors.

00:10:22: We tour robotics share data showing that their vision-only pipeline loses track of robot hand in twenty one point eight percent frames during a basic carrying task.

00:10:32: Wow!

00:10:33: They even recorded continuous dropout lasted four point three two seconds Over

00:10:37: four second of a robot flying totally blind while holding payload?

00:10:42: Why are cameras failing badly?

00:10:44: Vision occlusion, I mean when a robot reaches for an object its own arm often blocks the camera's view of its fingers.

00:10:51: or, you know...the lighting shifts in a warehouse where there is glare.

00:10:54: Vision-only systems just fail under dynamic

00:10:57: conditions."

00:10:57: Which completely explains that SEMG respans.

00:11:00: Weir actually had to start using electromyography to bridge that visual gap?

00:11:04: Right!

00:11:05: So they put an SEM G band on human teleoperator who was driving robot.

00:11:10: Because the camera can't see what the robot's fingers are doing, The band reads electrical impulses in human operator forearm muscles.

00:11:16: That

00:11:17: is crazy!

00:11:18: It tells a system exactly how hard hand is gripping completely bypassing need for a camera.

00:11:23: It honestly reminds me of a Formula One race car, like you can build the most powerful AI engine in the world.

00:11:30: But if the rubber of the tires fails under load?

00:11:32: You're just crashing into the wall.

00:11:33: yeah The hands are literally the tires of the humanoid robot and what's crazy is that the industry Is completely fractured on how to actually build them.

00:11:42: Figure as Brett Adcock publicly stated that using tendon-based hands was his biggest engineering mistake.

00:11:48: Right!

00:11:49: But then, one ex says Burnt Bornich stepped up and aggressively defended the exact same tendon architecture.

00:11:55: Well

00:11:56: because tendons give you incredible dexterity kind of like a human hand but they snap under heavy industrial loads.

00:12:02: Motorized joints gives us strength but their bulky and lack precision.

00:12:06: They just haven't solved mechanics yet.

00:12:08: So hardware is basically bottlenecked.

00:12:10: It's exactly why this software side is sprinting to compensate.

00:12:14: Look at the OM-One model from Zipank Fuze team, oh yeah I saw that it has zero shot generalizing across different robotic arms and humanoids using a seven degree of freedom on the body hand.

00:12:26: Yeah And they're doing with Zero Telly operation.

00:12:29: Wait, let me stop you there.

00:12:30: Zero shot generalizing means it can look at an entirely new object that has never encountered in training and just figure out how to pick up on the first try?

00:12:38: Yes!

00:12:39: There's no human driving with a VR headset.

00:12:44: It learns directly from raw human manipulation data.

00:12:46: That's

00:12:47: incredible.

00:12:48: And it goes further than that.

00:12:49: the Nvidia gear Carnegie Mellon and UC Berkeley Empire system is actually letting AI coding agents refine and rewrite robot policy code Directly on the physical hardware in a closed loop.

00:13:01: The AI is rewriting its own physical instructions in real time.

00:13:05: Yes, and they hit a ninety-nine percent pass rate across four real world tasks which even included inserting a GPU and securing zip ties.

00:13:13: Zip Ties?

00:13:13: I mean that's a nightmare for my human fingers let alone robot!

00:13:16: Right...and the data efficiency just skyrocketing there this diffusion policy called Refine DP.

00:13:23: Think of Diffusion Policy as like noise cancellation from movement.

00:13:26: it takes messy jerky actions to smooth them out into precise task.

00:13:29: Okay Got It.

00:13:30: Well, they achieved a ninety-five to ninety seven percent success rate using only fifty teleoperated demonstrations.

00:13:36: Wow!

00:13:37: Just fifty?

00:13:38: Yeah that is the twentyfold gain in efficiency.

00:13:41: and on the day it shipped an open source evaluation found GPT six Astra scored ninety five percent under robot control task compared just forty percent.

00:13:49: for fable five point one plus astra.

00:13:52: use six point two times fewer output tokens do it.

00:13:55: so this software is quite literally dragging hardware into future.

00:13:59: but You know, if the U.S and China are locked in this massive battle over software compute... ...and these huge data factories what is happening in Europe?

00:14:07: Well, Europe's strategy is a complete pivot from all that hype.

00:14:11: They aren't really chasing the sci-fi humanoid dream.

00:14:14: they're leaning hard into practical industrial physical AI

00:14:17: Which makes sense for them.

00:14:18: Yeah, the ecosystem there now has over a dozen companies building complete robots.

00:14:22: You've got Wandercraft, PL Robotics, Engineered Arts, Agile Robots and we just saw Seco, NERI Robotics in Qualcomm partner up to build physical AI hardware locally

00:14:32: And Nuriari also acquired at Latest Robotics recently amidst a much wider tech funding wave across the continent

00:14:40: right?

00:14:40: Exactly They are treating it strictly as an industrial deployment challenge.

00:14:46: IMEC over in Belgium just convened an advisory board to frame the entire robotic space, so power sensing latency strictly as a semiconductor problem.

00:14:57: Because yeah when you put server-level compute inside of mobile battery powered chassis heat and power draw become your absolute physical limits.

00:15:04: if they can't just slap massive liquid cooling tower on back walking robot You have solve thermal dynamics at chip level And the inaugural European Physical AI Summit in Munich proved this mindset perfectly, I think.

00:15:17: Oh absolutely!

00:15:18: They brought together defense logistics and automotive leaders to talk strictly about what actually works on the factory floor today not in ten years

00:15:25: Right but speaking of defense that brings up the starkest sector split in the entire industry right now.

00:15:30: Yeah...this part is fascinating.

00:15:32: Investor Shane Neiman pointed out that physical AIs basically fracturing into two distinct camps using the exact same technology stack.

00:15:40: You have a well-funded defense camp building machines to fight, and then massive often overlooked health care Camp Building Machines to keep people alive.

00:15:49: And

00:15:49: the stakes in healthcare are just astronomical!

00:15:51: Just look at the capital moving around... Medtronic recently paid seven hundred million dollars to Cornerstone Robotics just for the distribution rights of their entire surgical robot outside US.

00:16:03: Just for distribution right?

00:16:05: Right, why would they drop almost a billion dollars on distribution?

00:16:08: because Medtronics spent thirteen years and thousands of surgeon sessions building their own Hugo platform from scratch a project they started all the way back in twenty twelve as Project Einstein.

00:16:19: Building precision hardware that can pass medical regulations takes decades.

00:16:23: Medtronic basically decided it was cheaper to just spend seven hundred million dollars than to repeat that thirteen year build time.

00:16:29: Yeah,

00:16:30: The barrier to entry is massive

00:16:31: It Is.

00:16:32: But the surgical robotics market is projected to jump from roughly seventeen billion dollars in twenty-twenty six To over forty five billion by twenty thirty three

00:16:41: And the capabilities they actually have right now sound like pure science fiction.

00:16:45: During this same exact window, we're looking at a surgeon in Rome performed alive robotic prostate procedure on a patient in Beijing.

00:16:53: It's mind-blowing!

00:16:54: The robotic arms in Beijing were reproducing the surgeons' exact micro movements In real time over five G and fiber optic networks eight thousand kilometers away.

00:17:05: That honestly perfectly brings this whole thing full circle because we started out talking about telecom infrastructure and fiber optics driving these massive wall street valuations for human or robots, And we end with those exact same fiber optics literally enabling a surgeon to save a life from across the planet.

00:17:20: today

00:17:20: It's great perspective Yeah

00:17:21: Which leaves us with final thought as you evaluate technology coming into your own operations.

00:17:27: If Amazon needs thirty percent more reliability in maintenance staff For their next generation automated facility and China is building tens of thousands of humanoids just to harvest failure data, Is the smart manufacturing facility tomorrow actually less about autonomous hands-off labor?

00:17:44: And more about managing a massive steel wrapped data harvester machine that constantly needs human intervention To be fixed?

00:17:52: it completely changes.

00:17:53: The ROI calculation of the future

00:17:55: It really does.

00:17:56: if you enjoyed this episode new episodes drop every two weeks.

00:17:59: also check out our other additions on Smart Manufacturing in industrial AI, digital construction and connected tools.

00:18:05: Thanks so much for joining us on this deep dive.

00:18:07: don't forget to hit subscribe.

00:18:08: we'll see you next time!