Best of LinkedIn: Robotics CW 34/ 35

00:00:00: provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about robotics in calendar weeks thirty-four and thirty five.

00:00:09: Frenness is a B to B market research company that supports robotics and smart manufacturing providers with building feature by future competitive intelligence.

00:00:17: That shows exactly how their product stacks up against the competition.

00:00:20: you can find more info in the description

00:00:22: And welcome into the deep dive.

00:00:24: I mean, if you're a professional in the smart build and manufacturing space this conversation is tailored specifically for You today.

00:00:31: yeah absolutely we are completely bypassing The fluff and the hype cycle.

00:00:36: We're focusing strictly on the bleeding edge of what's actually working In the field right now based On our source material.

00:00:42: because i mean If you imagine watching A five minute video someone say potting a plant and then immediately knowing how to do it perfectly using the mechanical body you've never operated before.

00:00:52: Which

00:00:52: is wild even think

00:00:53: about!

00:00:54: Right, for humans observational learning is completely normal.

00:00:57: but for a robot translating a two-D video into three D physical movement well that used to be impossible until right now.

00:01:05: So today we're looking at how foundation models are moving off screens into the physical world.

00:01:09: We'll look at humanoid scaling, medical robotics and uh...the kind of unglamorous reality of industrial deployment.

00:01:17: Yeah it's a massive paradigm shift.

00:01:19: I mean for decades The standard was rigid code You program a robotic arm to move exactly ten centimeters left Right If the environment changes, if a part is bumped just two millimeters.

00:01:30: The robot fails.

00:01:31: it had no real spatial awareness

00:01:33: which was so fragile extremely

00:01:35: fragile.

00:01:35: but that bottleneck is shattering and I think the most compelling proof of this from our sources is skilled AI's new S one model.

00:01:44: yeah i saw the insights on this from Garfkana Deepak Pathok and Ryan Embry Jones.

00:01:48: they were talking about s-one using in context learning for the physical world right?

00:01:53: But I kind of need to pause here because text is a terrible way to describe physics.

00:01:58: You can't just type a prompt into a robot, explaining the exact torque and friction required.

00:02:15: video prompt.

00:02:16: Meaning like a first-person perspective?

00:02:18: Exactly!

00:02:18: A video shot from the human's point of view, it'll just show the model—a video —of a human doing the task.

00:02:24: The AI holds that video in its context window analyzes movements extracts underlying intent and dynamically maps to its own mechanical joints.

00:02:33: Wait I have challenge though.

00:02:35: How does an AI bridge gap between human anatomy and robotic hardware?

00:02:41: Well, I mean a robot arm doesn't have the same muscle groupings or degrees of freedom as A human arm.

00:02:47: and that's exactly The genius of the foundation model approach.

00:02:51: The model has already been trained on massive amounts Of generalized physical data.

00:02:55: it knows how physics works?

00:02:56: How gravity behaves right.

00:02:57: so when It watches the human video it isn't trying to copy the Human as elbow angle degree by Degree.

00:03:03: it understands the goal getting dirt into a pot for example And calculates how To achieve That goal using its own unique kinematic structure.

00:03:10: oh wow And the benchmarks they hit doing this are honestly staggering.

00:03:13: They really are!

00:03:15: They tested S-One on completely unseen, long horizon tasks like a tenement process of potting a plant.

00:03:21: Traditional vision language action models usually hit about a nine percent success rate on multi step tasks like that.

00:03:27: Yeah...nine percent is basically useless in factory

00:03:30: Right but skilled AISS one hit a sixty six per cent success rate.

00:03:35: and the

00:03:35: timeline what should grab your attention if you're listening or working manufacturing?

00:03:40: Ryan Embry-Jones pointed out an experiment where a human demo was recorded at nine point two two p.m.. Okay by nine point to seven p.M The S one model began autonomous execution.

00:03:51: wait really five minutes?

00:03:52: Five minutes from a single video demonstration To A robot performing, a complex unseen physical task.

00:03:59: and it demonstrated actual common sense too in the training Video.

00:04:02: the Human used a watering can But when it was the robot's turn, there were only a cup of water available.

00:04:07: Right and didn't just freeze up?

00:04:09: Exactly!

00:04:09: The robot didn't crash because the watering can was missing.

00:04:13: It adapted grab-the-cup and poured in.

00:04:15: That kind visual reasoning is popping everywhere now.

00:04:18: Marcel Minge highlighted an amazing experiment with DeepSeek's new V four flash vision model.

00:04:25: Oh yeah, the humanoid experiment Yeah

00:04:27: he plugged this vision model directly into camera feed of a humanoid named Ralcox and he told it to safely approach a person sitting at the laptop.

00:04:36: But instead of just spitting out basic image description, like I see a table... The model actively reasoned about this space.

00:04:43: Right!

00:04:43: It mapped it out.

00:04:44: It identified that table with some scattered items specifically as navigation constraints.

00:04:49: It plotted a path And this is part really surprised me.

00:04:52: It explicitly acknowledged its own uncertainty Which was huge!

00:04:55: Yeah..it told user- from this single, two-D image.

00:05:04: Which is a massive leap forward!

00:05:06: True autonomous reasoning isn't just about knowing what to do.

00:05:09: it's about knowing that you don't know.

00:05:11: Right evaluating risk.

00:05:12: Exactly in Bayesian terms It tats its probability.

00:05:15: the system processes perception establishes contest evaluates its own uncertainty and then decides on safe action.

00:05:22: And That brings up an interesting strategic divide if You have this incredibly smart software how Do you pair with hardware?

00:05:30: Leo Sue pointed out that Anthropic recently introduced an open model hardware standard.

00:05:35: Yes So instead of building a robot body and trying to cram a brain into it, Anthropic seems to be building a massive brain That is actively looking for bodies to pilot like a universal driver For any vehicle?

00:05:48: That analogy is spot-on.

00:05:50: Historically robotics has been very bottom up.

00:05:52: you build the motors design The chassis and then try to layer intelligence on top.

00:05:56: Yeah.

00:05:57: Anthropic is taking a top-down approach, they're taking their massive advantage in generalized reasoning with Claude and pushing that intelligence down into the physical world standardizing how hardware should talk to their brain.

00:06:09: But I mean here's the problem with universal brain strategy.

00:06:12: it doesn't matter How smart the AI Is if you can't manufacture The physical bodies at scale which leads us right Into the brutal reality of humanoid Hardware.

00:06:22: Tim Carden broke Down some numbers showing A massive disconnect In the market Right now.

00:06:26: Humanoid startups have raised eight point six billion dollars this year, but the valuations are completely detached from actual production capacity.

00:06:34: The contrast is jarring.

00:06:36: honestly look at Agibut.

00:06:38: there a Chinese manufacturer currently shipping forty four percent of the world's humanoids.

00:06:42: Wow

00:06:43: almost half.

00:06:44: Yeah, they actually have supply chains and factory output.

00:06:47: Yet their valuation is roughly one sixteenth of figure which an American company that has phenomenal AI partnerships but there are essentially still running pilot programs.

00:06:57: They haven't scaled mass deliveries.

00:06:59: So the market is clearly placing a massive premium on software potential over current manufacturing capacity.

00:07:05: Investors basically betting.

00:07:07: if figure solves general intelligence The hardware just becomes cheap interchangeable commodity.

00:07:13: But that might be a fatal miscalculation by investors.

00:07:15: If you connect this to the bigger picture of heavy manufacturing, building reliable humanoid at scale is an incredibly deep moat.

00:07:23: Oh definitely!

00:07:24: Meet Gultikin pointed out that electric vehicle makers are entering the humanoid space on an industrial scale.

00:07:30: We're talking companies like Tesla X-Pen BYD and Sherry rapidly spinning up humanoid production lines.

00:07:37: Well, because an EV is basically just a giant robot on wheels anyway.

00:07:40: Right?

00:07:40: Exactly!

00:07:41: These car companies already have the entire infrastructure.

00:07:44: they aren't starting from scratch.

00:07:46: They're reusing their existing technology stacks for high torque motors advanced battery cells mass manufacturing

00:07:53: And Srinivasa Rao Alluri argued a crucial point here.

00:07:57: for anyone listening who actually deploys hardware at an industrial scale the winner won't be The company with the smartest lab prototype right?

00:08:05: Scale is going to be decided by engineering consistency and failure data.

00:08:09: when you deploy thousands of robots You start dealing with the law of large numbers statistical variations in motor torque thermal resistance

00:08:17: or gearbox backlash.

00:08:18: For Anyone Who hasn't had to Maintain Heavy Machinery Gearbox Backlash Is that microscopic bit of play or looseness between the teeth and gears.

00:08:26: In a lab demo, it's tiny wobble.

00:08:29: but on factory floor running twenty four hours per day that microscopic wobble turns into violent vibration.

00:08:36: It causes thermal de-erating.

00:08:37: where the motors get so hot they actually lose torque And eventually robot literally tears its own joints apart.

00:08:43: Which

00:08:43: brings us to two intense hardware debates quietly holding entire humanoid industry back.

00:08:50: The first is about hands.

00:08:53: Christian Rake highlighted a massive divide in how engineers are designing robotic hands for the real

00:08:58: world.

00:08:58: Tendon-driven architectures versus in finger actuation, so tendon driven hands place heavy motors up into robot's forearm and they use synthetic cables pulling through wrist to move fingers just like human tendons

00:09:11: Exactly!

00:09:11: They're incredibly lightweight and highly dexterous.

00:09:14: But there an absolute nightmare to maintain.

00:09:16: I mean if single tendon snaps on factory floor you can't fix it.

00:09:19: You often have disassemble entire arm

00:09:21: And on the other side, you have in-finger actuation.

00:09:25: This puts tiny motors directly inside the palm or the knuckles.

00:09:29: From a maintenance perspective it's brilliant.

00:09:32: If a finger breaks A text just unplugs and snaps an new one In But

00:09:36: trade off.

00:09:36: there is severe.

00:09:37: It makes hands heavy Clunky creates massive thermal management problem.

00:09:42: Those tiny motors get incredibly hot trapped Inside fingers.

00:09:46: The industry currently split.

00:09:48: Nobody knows which design will win out for mass production.

00:09:51: The second major bottleneck is even simpler, honestly.

00:09:54: Power!

00:09:55: Dale Tutt issued a stark warning in his post arguing that battery run time charging speed and energy density not AI models are the real constraints keeping humanoids out of factories right now

00:10:05: because it all comes down to unit economics.

00:10:07: if a humanoid has its utility drops to zero.

00:10:12: If it only runs for three hours and then spends four hours charging, the financial argument for deploying it completely collapses.

00:10:18: And here is the catch-twenty two.

00:10:19: as these visual reasoning AI models get smarter they require significantly more onboard compute power which drains the battery even faster.

00:10:28: By the way if you want keep tracking how this hardware bottlenecks resolve make sure hit subscribe so that we can catch our next deep dive.

00:10:34: Yeah, it's definitely a space to watch.

00:10:36: But you know while humanoids are still struggling to figure out battery life for factory floors there is an entirely different sector of robotics where raw power doesn't matter at all right?

00:10:47: A space where millimeter level precision and life or death.

00:10:50: reliability Are the only metrics that count.

00:10:53: You're talking about medical and surgical robotics.

00:10:55: Absolutely.

00:10:56: Mayuri Lomba shared projections That this market alone will hit twenty seven point one four billion dollars by twenty thirty.

00:11:04: Well, what stood out to me from the sources is how nature of this competition has shifted.

00:11:09: It's not just about a company rolling in the flashiest piece of futuristic tech anymore...

00:11:13: No it's entirely about workflow integration!

00:11:16: Yay Elle brought up a brilliant point regarding Johnson & Johnson's OTT system which got FDA authorization.

00:11:23: The battleground now operates room efficiency.

00:11:26: Hospitals don't buy a robot because they have cool AI.

00:11:30: They buy if it seamlessly fits their existing work flow turns over operating rooms faster and makes advanced technology feel ordinary on a busy Tuesday morning.

00:11:39: That makes total sense!

00:11:40: I mean, a surgeon isn't going to use the tool that takes two hours to set up but at the same time...the pure capability of these systems is mind-bending.

00:11:48: It really is.

00:11:49: Alex Falstive shared video about a teleoperated microsurgical assistant who looks like CGI.

00:11:56: It was suturing a single kernel of corn using this ultra fine medical thread, treating the corn exactly like a submillimeter human blood vessel.

00:12:05: The mechanics behind that are fascinating.

00:12:07: The robot uses tremor filtering?

00:12:09: Oh, to stop the shaking... Yeah!

00:12:10: A human hand naturally shakes a few micrometers no matter how skilled the surgeon is.

00:12:15: The robotic system mechanically filters out those micro-tremors scaling down the surgeon's gross hand movements into microscopic perfectly stable actions Things

00:12:23: that're physically impossible for humans to achieve alone.

00:12:26: Exactly, it even has automatic tool changing so the system can swap out micro forceps for micros scissors in seconds.

00:12:33: It effectively allows a general surgeon to perform at level of an elite micro surgeon And this isn't just happening in test labs either.

00:12:40: Professor Dr Yan Brido posted about performing the very first European spine surgery using Medtronic's stealth AXIS platform in Cologne.

00:12:49: That

00:12:49: is a huge milestone!

00:12:51: Yeah,

00:12:51: and his main takeaway was that spatial navigation and robotic action are finally merging into one workflow... The robot isn't some clunky, separate step anymore.

00:13:00: It guides the tools based on a three-D map of patient R spine in real time.

00:13:04: And we're also seeing new physical forms that completely challenge our assumptions about medical robots.

00:13:10: Stefano M. Cinecropi highlighted project out UC San Diego involving a robot nicknamed Surgie.

00:13:16: Oh this one is wild!

00:13:17: Surgies is a sixty pound five foot tall teleoperated humanoid robot and it actually assisted with a real live gallbladder removal.

00:13:25: Right, which is incredible because most surgical robots are these massive eighteen hundred pound machines that literally require a hospital to reinforce their floors.

00:13:33: Yes you have to rebuild the operating room just to fit them but surgery is human-sized.

00:13:39: it fits wherever a human nurse would stand.

00:13:42: The access implications there are huge.

00:13:44: if you can deploy a lightweight human size surgical robot You suddenly bring specialist care to rural communities disaster zones remote field hospitals Places where flying and a specialist just isn't possible in time.

00:13:57: And if five-foot humanoid is wild enough for you, honor Sezgin shared research on the caterpillar inspired soft magnetic milli robot developed at The University of Hong Kong.

00:14:07: Yeah We are talking about complete opposite end of the size spectrum here.

00:14:11: Yeah, totally different approach.

00:14:13: This Millie robot doesn't have internal motors or batteries.

00:14:16: it is made of soft materials embedded with magnetic particles.

00:14:20: So how does it move?

00:14:21: An external magnetic field is used to push and pull those particles.

00:14:25: That allows the robot to crawl across wet slippery biological tissue inside the human body.

00:14:30: It can carry upto a hundred times its own weight.

00:14:32: that's

00:14:32: unbelievable.

00:14:33: The ultimate goal is highly targeted localized drug delivery without requiring any invasive surgery.

00:14:38: It is incredible to just sit back and marvel at that contrast.

00:14:42: I mean, we're talking about microscopic robots... crawling through human tissue on one hand and then the other, massive AI foundation models preparing to control fifty-ton heavy machinery out in mud.

00:14:54: Which

00:14:54: brings us into this final and arguably most difficult piece of puzzle industrial deployment operations.

00:15:00: because taking these physical AI models from a sterile lab putting them to work is where most robotics companies fail.

00:15:14: Caterpillar has formed a strategic partnership with Field AI to bring foundation models directly into highly unpredictable, heavy industry jobsites.

00:15:22: Which is huge test for the technology.

00:15:24: Huge!

00:15:25: But deploying AI in that kind of environment exposes an unglamorous reality.

00:15:29: Patrick Walsh articulated perfectly He noted In a lab failure was just bad demo video but in field it's massive revenue leak.

00:15:38: Right, let's walk through what that actually means for a business.

00:15:42: If a robotic cell goes down at customer site the robotics provider usually has to fly out two specialized engineers three days just observe and reproduce failure which

00:15:54: is incredibly expensive.

00:15:55: then software team pushes in over-the air patch.

00:15:59: But if the company doesn't have proper regression testing, that patch is a massive liability.

00:16:05: And for context, regression-testing as standard software practice where you prove your new bug fix didn't accidentally break something else.

00:16:12: It's like fixing a leaky pipe in house but accidentally shutting off water heater.

00:16:17: If you don't run regression tests You might fix robot navigation issue but accidently break its ability to pick up heavy box.

00:16:24: And when that happens on a live site, you destroy the customer's trust.

00:16:27: The procurement department demands another four weeks of supervised operation and Walsh's core argument is that procurement departments do not buy lab capabilities.

00:16:37: No, they buy predictability

00:16:39: Exactly!

00:16:40: They want a vendor to walk in and say here are four hundred previously failed scenarios That our robot now passes.

00:16:47: And Here Is the exact simulation proof.

00:16:50: It's crazy because traditional software engineering solved regression testing twenty years ago.

00:16:54: The robotics industry still treats it like luxury... ...and they're bleeding cash In lost renewables Because of it.

00:17:00: It

00:17:00: really Are.

00:17:01: So if you are a facility manager listening to this, what does a mature deployment stack actually look like?

00:17:07: Well Shreya Lepperloo provided an excellent highly structured framework based on what they're building at AWS.

00:17:12: Okay The AI model itself the brain is just one piece of the puzzle.

00:17:17: Real deployment requires four interconnected layers.

00:17:20: first is physics grounded data code and text are not enough.

00:17:24: The training data has to account for real-world friction, harsh lighting changes dynamic contact

00:17:30: forces.

00:17:30: And the second layer is digital twin simulation.

00:17:32: right?

00:17:33: Training an AI by having a physical robot fail repeatedly in the real world.

00:17:37: it's just too slow...too expensive and wildly dangerous

00:17:40: Absolutely.

00:17:41: You have to build a digital twin, essentially a hyper-realistic video game physics engine where the robot can crash a million times per second safely stress testing edge cases before it ever gets near human worker.

00:17:54: Okay what's that third layer?

00:17:55: Third is the continuous cloud to Edge loop.

00:17:58: you deploy a compressed smaller AI model locally on the robot itself so it can react instantly without latency.

00:18:06: but All of its real-world edge cases must continuously flow back up to the massive cloud model, To improve the entire fleet.

00:18:13: And fourth layer is Fleet Operations at scale Which means robust version control Secure rollbacks when a patch fails and total observability Of hardware

00:18:21: help.

00:18:22: Exactly

00:18:23: So hearing all that it sounds like the actual bottleneck For getting autonomous robots onto factory floors.

00:18:28: Isn't even AI algorithms anymore?

00:18:30: The tech's basically there?

00:18:32: You're right It isn't math.

00:18:33: Moritz Frender gave a fantastic reality check on this regarding Wandelbotz Industrial AI Cloud.

00:18:39: He noted that when you look at automotive and heavy manufacturing, the technical AI ambition is rarely the holdup.

00:18:45: The real bottleneck is almost entirely organizational

00:18:48: Which means office politics?

00:18:50: Budget fights And risk management?

00:18:52: Essentially yes It's a management problem.

00:18:54: Who actually owns the budget for general purpose robot?

00:18:58: Does it come out of IT or operations Or facilities?

00:19:01: And who was willing to sign their name and take on the personal career risk of deploying a fully autonomous system.

00:19:07: Right,

00:19:07: And most importantly what does that daily operating model look like?

00:19:11: Organizations literally have to rewire how human workers operate to accommodate their robot.

00:19:17: Navigating Human Change Management is significantly harder than writing better neural network.

00:19:22: When you step back at all this we've covered massive ground today.

00:19:26: I mean, we've gone from AI learning how to pot plants from a single five minute video.

00:19:30: two humanoids waiting on better thermal management too tremor filtering microsurgery all the heavy machinery deployments stalling out because of corporate budget disputes.

00:19:39: where does this leave

00:19:40: us?

00:19:41: It leaves us with a highly provocative realization, I think about the future of this industry.

00:19:46: If the real bottlenecks to scale are organizational risk and unpredictable field failures then The ultimate winner in robotics will not be the company With the flashiest AI or the most human like chassis right?

00:19:59: The company that dominates the market.

00:20:00: Will we the one whose robots can silently detect troubleshoot, and compensate for their own mechanical drifts in gearbox backlash before a human operator ever realized as if failure was about to happen.

00:20:12: Predictive maintenance handled autonomously at the hardware level

00:20:30: is.