Best of LinkedIn: Smart Manufacturing CW 13/ 14

00:00:00: This episode is provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about smart manufacturing in calendar weeks thirteen and fourteen.

00:00:09: Frenness is a B to B market research company that supports Smart Manufacturing providers with building future-by-feature competitive intelligence That shows exactly how their product stacks up against the competition.

00:00:20: You can find more info in the description.

00:00:23: So look Imagine you spend four weeks building this massive, highly complex physical assembly line.

00:00:29: Oh yeah the classic nightmare

00:00:31: right?

00:00:32: You've bolted down all the robots...you run the power..You are ready for day one and in a very moment when turn it on two robotic arms just swing into exactly same space and violently collide.

00:00:42: Damn

00:00:43: millions of dollars in damage instantly not to mention the week's delays

00:00:47: Exactly!

00:00:48: And for decades that risk was well the brutal cost doing business and manufacturing.

00:00:54: It really was an accepted reality.

00:00:56: I mean, you designed it on a screen You built in physical space and you just prayed.

00:00:59: the translation worked.

00:01:01: But if you are stepping onto a factory floor today or even looking at this strategic roadmaps for next few years you're witnessing a fundamental rewrite of that whole process.

00:01:12: Absolutely and it's exactly what we are getting into today, We doing deep dive in to the top smart manufacturing trends across LinkedIn over past two weeks

00:01:23: Right were moving away from just digital ambition.

00:01:27: You know isolated pilot programs The flashy but totally useless dashboards Were removing an era of ruthless scalable execution

00:01:35: And we are going to cut right through the noise, and give you a crash course in how the factory floor is fundamentally changing.

00:01:41: But before I get to talk about really flashy stuff like humanoid robots or agentic AI making on-the-fly decisions... We have to establish the ground reality!

00:01:52: Exactly.

00:01:53: The plumbing because across all the sources we analyzed, the biggest bottleneck to smart manufacturing isn't a lack of ambition it's just fragmented chaotic data.

00:02:03: you simply cannot build a smart autonomous factory on bad

00:02:07: plumbing.

00:02:07: No You can't!

00:02:09: The clearest barrier to scaling any kind of intelligent manufacturing right now is disconnected data.

00:02:14: I mean if... If you walk into typical facility and have your PLCs the programmable logic controllers which are basically firing the actual motors and valves.

00:02:23: Right, a physical layer?

00:02:24: Yeah!

00:02:25: And then you have SCADA systems—the screens trying to monitor those processes —and above that, the MES... ...trying to track overall production orders….

00:02:31: …And none of these systems naturally talk with each other.

00:02:34: They just live in total silos

00:02:35: Which actually brings us into really counter-intuitive insight from Jen's Lamparth over at Siemens.

00:02:40: He brought up this fascinating concept of AI correlation trap.

00:02:43: Oh I love his point.

00:02:44: It makes so much sense when we think about it

00:02:47: Because normally if you manage a facility Having a machine run perfectly smoothly for six months is the ultimate goal.

00:02:53: You celebrate that, but Lamparth points out that those perfectly running machines generate vast amounts of highly repetitive data which

00:03:01: Is great for your daily uptime.

00:03:03: But it is absolutely terrible for training AI

00:03:05: model exactly because the AI essentially learns from variance.

00:03:10: Yeah It requires anomalies.

00:03:11: It requires failures and it needs data on how the system actually recovered from those failures.

00:03:17: If you feed an industrial AI model a billion data points of a robotic arm doing the exact same, perfect motion every single day it doesn't learn to predict motor burning out

00:03:28: Right!

00:03:28: It just learns the mathematical signature what normal looks like.

00:03:32: Exactly

00:03:32: So when rare catastrophic failure event happens The AI is completely blind to it.

00:03:38: It just hasn't been trained on the complex data of industrial anomalies.

00:03:41: Wait, so just to be clear because the machines are running perfectly.

00:03:44: ninety-nine percent at a time... ...the AIs essentially just memorizing what a good day looks like.

00:03:49: Yep!

00:03:50: Its completely useless in crisis Because its never actually seen one.

00:03:54: That's the correlation trap In a nutshell.

00:03:56: Wow That is a massive roadblock.

00:03:58: It really is.

00:03:59: to get around it, you need AI that actually understands the physical engineering context not just generic data points.

00:04:05: and that's why Siemens is currently investing over one billion euros into an industrial foundation model.

00:04:11: right now

00:04:12: Just to natively understand industrial languages?

00:04:15: Right

00:04:15: exactly they aren't just buying off-the-shelf AI.

00:04:17: we're talking about models.

00:04:19: They can natively read three decad files parse raw sensor data And actually understand PLC code.

00:04:25: But if we pull back and look at how that data is structured, Aaron Luck articulated another massive headache in his proposal for OpenUSD.

00:04:33: He pointed out the nightmare of object identity fragmentation.

00:04:36: Ooh!

00:04:36: The naming conventions… it's a mess.

00:04:38: It IS because even you have advanced AI... ...the physical data systems fundamentally cannot agree on what to call single-physical objects on your factory floor.

00:04:48: Right Consider just a single physical bracket on an assembly line.

00:04:52: The architect who designed this building might label it with a specific Revit element ID.

00:04:57: Then the PLM software that tracks the engineering specs calls it by unique part number,

00:05:03: and then the factory floor software talking to machines call it an OPC UA.

00:05:08: noted

00:05:08: Exactly!

00:05:09: It's exactly same piece of metal but systems are effectively speaking completely different languages.

00:05:15: So applying AI into a factory right now is I mean, it's like giving a brilliant scholar a massive library of books but every single book is written in language they can't even read.

00:05:25: That is the perfect way to visualize that.

00:05:27: The Scholar has all processing power in this world and the underlying text is just noise for them.

00:05:32: Yeah!

00:05:32: And Aaron Luck's proposal for the Alliance For OpenUSD is trying frame this as separation from concerns – standardizing how we actually link object identities across these systems inherently disagree, because until you fix that fundamental data infrastructure the AI simply cannot scale.

00:05:50: It'll just hallucinate or completely fail to connect with dots.

00:05:53: but that identity fragmentation isn't just an IT headache it becomes a massive financial liability at this.

00:05:59: second external pressures hit.

00:06:01: Jacobo Loretka saw provided really stark reality check on this.

00:06:05: He did!

00:06:05: he noted for the past decade industry four point oh was sold as growth story You know?

00:06:11: A way get competitive edge

00:06:12: Right, but now in twenty-twenty six he argues it strictly about survival.

00:06:17: He's seeing factories across Europe just bleeding margin because they lack this real time data infrastructure.

00:06:24: Yeah

00:06:25: when a sudden energy price spike hits the market that lack of connected data becomes a full blown crisis.

00:06:30: yeah if your systems are siloed you don't have realtime consumption data mapped to your production output

00:06:35: so You literally don't know where your energy losses are until what?

00:06:39: The spreadsheet arrives from accounting three weeks later.

00:06:41: yep

00:06:41: And by then, the margin on everything you built that month is already completely wiped out.

00:06:46: Casol frames this not as an IT problem but a board-level risk

00:06:50: Because manufacturers who actually invested in robust operational backbone can respond to market and energy shocks for hours adjusting their production dynamically

00:06:59: While those who didn't are just doing damage control.

00:07:01: Okay so let's look at companies that did invest.

00:07:03: Let's say you fix your plumbing, Your data is standardized and your systems are finally speaking the same language.

00:07:09: The

00:07:09: dream scenario?

00:07:10: Right!

00:07:10: You take that clean data And build a mirror of the physical world.

00:07:14: This moves us perfectly into our second theme Digital Twins and Simulation-led Manufacturing.

00:07:20: Once your data foundation Is actually solid... ...the reliance on Physical trial & error completely vanishes.

00:07:27: You don't need to test process changes On physical steel anymore.

00:07:30: You test them in the digital twin

00:07:32: Exactly.

00:07:33: You test the physics, timing and material flow all inside of the twin.

00:07:38: It is a massive shift from reactive troubleshooting to predictive validation.

00:07:42: Sean Cia shared a fascinating breakdown of BMW's iFactory that completely changes the timeline for how we build cars.

00:07:49: They built this Twin on NVIDIA Omniverse.

00:07:52: Oh!

00:07:52: The scale in this is just incredible...

00:07:54: It really is.

00:07:55: To put it into perspective A new car model that used to take four weeks to physically validate on a real assembly line is now collision checked and fully validated inside of digital twin in just three days.

00:08:09: Three days, And they achieved a virtual start-of production more than two years before the actual physical operations that their plant even began.

00:08:18: The

00:08:19: physics of torque and gravity, steel haven't changed but management discipline has.

00:08:24: they are running all there highly expensive high stakes experiments virtually

00:08:28: Yeah.

00:08:28: if a robot arm is going to swing too wide hit safety barrier They see it on screen two years before barriers manufactured.

00:08:35: It's like we're finally treating factory floors, like software engineers treat code.

00:08:39: I mean you would never push untested code to a live commercial website.

00:08:43: You run it in a sandbox first

00:08:45: Right!

00:08:45: You test it thoroughly.

00:08:46: So why on earth have we spent decades pushing untested workflows To alive multi-million dollar assembly line?

00:08:53: You wouldn't Not if you had a choice.

00:08:55: And that mindset is exactly what Jerry Keen highlighted regarding PepsiCo.

00:09:00: PepsiCo was using Siemens Digital Twin Composer, integrated with NVIDIA Omniverse to just completely revamp how they modernize their facilities... ...and Keen

00:09:08: makes a crucial point here this isn't

00:09:11: a science project!

00:09:12: No not at all.

00:09:13: They aren't paying engineers to build custom three-D environments.

00:09:17: Just show off executives in the boardroom.

00:09:19: There are using out of box platforms To simulate real material flow.

00:09:24: But how does that material flow simulation actually translate to the bottom line?

00:09:28: Like, what are their real-world results.

00:09:30: It runs physics and timing engines to spot bottlenecks before they ever buy equipment so They can model thousands of bottles moving down online And see exactly where a jam will occur if they increase this speed by say five percent.

00:09:42: Wow Because of this PepsiCo has seen up to twenty percent throughput games and there catching ninety percent Of potential issues Before a single dollar of capital expenditure is even spent.

00:09:53: That's massive, catching ninety percent of issues virtually.

00:09:56: It is incredible to see that at an enterprise scale but what I find even more compelling how this trickles down into smaller operations.

00:10:03: Yeah David Morley brought us right down the earth!

00:10:05: He did.

00:10:05: he was talking about every day CNC shops.

00:10:09: If you've never stood in front of a five-axis CNC machine, You might not realize how terrifying the setup process can

00:10:15: actually be.

00:10:16: Oh it is intense!

00:10:17: You are programming hardened steel to move at incredibly high speeds toward a very expensive metal fixture

00:10:23: Right and if your toolpath is off by even a fraction of an inch The machine crashes And you just destroyed a tool worth tens of thousands Of dollars.

00:10:30: But Morley pointed out that digital twins allow these smaller shops To eliminate that costly trial & error too.

00:10:37: They simulate the exact toolpath, they check for collisions in the software and validate that whole process safely on a computer before the operator ever presses cycle start.

00:10:47: It is completely democratizing this simulation-led approach.

00:10:51: You don't need to be BMW to prevent catastrophic physical crash anymore.

00:10:56: Okay so we have simulated perfect factory And avoided all digital crashes But eventually you do press Cycle Start in real world.

00:11:06: Which brings us to our third theme, industrial AI and operational decisioning.

00:11:12: Because when the factory actually goes live you transition from predicting a clean simulated future To managing of very messy chaotic present

00:11:20: And The old way of managing that chaos is dying.

00:11:23: Brent Roberts outlined this transition beautifully.

00:11:26: He notes That model shifting toward outside in AI Where facility essentially orchestrates itself like one giant robot

00:11:34: An AI brain sitting on top of the operations software.

00:11:37: Exactly!

00:11:38: It consumes alive data from a digital twin, and it closes control loops in real time to eliminate production drift & scrap.

00:11:44: But what does closing those control loops actually look like for that person managing the floor?

00:11:49: It means end of traditional SATA and PLC lag.

00:11:53: In a traditional setup operators tune individual machine cells.

00:11:57: They literally wait for red light to flash onto screen.

00:12:00: But physical alarms travel slower than the actual physical process, right?

00:12:03: Yes.

00:12:04: By the time the human operator sees the alarm walks over and adjusts the dial... ...the machine has already produced fifty defective parts.

00:12:12: The new North Star metric for operations leads is shrinking the alarm.

00:12:16: to adjust time You have to move that metric from minutes or hours down to mere seconds.

00:12:22: And

00:12:22: Neymar Alassi gave a really fascinating example of this with Honeywell's Agentec AI, instead an operator staring at control room screen full blinking red lights just trying to dig through multiple data tags and figure out why the pump is feeling?

00:12:36: The AI acts as proactive partner.

00:12:39: Yeah it analyzes vibration temperature data until the operator says hey!

00:12:43: This asset will fail in four.

00:12:45: follow these exact steps to prevent it.

00:12:47: We're replacing reactive alarms with proactive step-by-step guidance,

00:12:51: which represents a profound shift in operational decisioning.

00:12:54: It turns the AI into an execution layer that actively stabilizes plant performance rather than just acting as a fancy dashboard That reports the bad news after the fact.

00:13:03: Okay But let me pause and push back on this for a second because if you manage A facility right now You might be rolling your eyes at The idea of An AI brain seamlessly orchestrating Your floor.

00:13:14: You know how messy a Tuesday afternoon actually gets?

00:13:16: Oh, for

00:13:17: sure.

00:13:17: The reality is always messier

00:13:19: And the data backs up that skepticism.

00:13:21: Michael Chu pointed out that McKinsey research shows sixty-one percent of these automation initiatives still completely miss their targets.

00:13:29: Fifty one percent It's huge number

00:13:31: If we have digital twins mapping everything and agentic AI predicting failures hours in advance.

00:13:37: Why there such massive disconnect between capability tech and actual results on shop floor?

00:13:43: The answer lies in how organizations are actually deploying it.

00:13:47: Both Victor M and Kudzai Mandatresa provided brilliant insights into this specific failure rate.

00:13:53: Companies fail because they apply yesterday's transformation logic to todays AI.

00:13:58: Management walks-in asks, what specific tasks can this AI automate?

00:14:03: Instead of asking where is the friction on our overall workflow?

00:14:06: So they're just taking a bad, inefficient process and using advanced AI to do that bad process faster.

00:14:14: They are as Victor M put it industrializing waste... Industrializing

00:14:18: waste?

00:14:18: That is the core issue!

00:14:20: If you don't step back and redesign workflow from blank sheet of paper automation amplifies your existing bottlenecks.

00:14:26: You get a faster traffic jam

00:14:28: Exactly.

00:14:29: But could I manage raise a hint on something even deeper regarding human element The builders of manufacturing AI often treat factories like generic enterprise software environments.

00:14:39: They completely forget that critical operational logic lives exclusively in the operator's heads, precisely Think about how many factory workflows are held together by the operator who just knows That a specific machine needs to be recalibrated slightly when the ambient humidity in the room changes or

00:14:58: The operator who can literally hear a faint change in the pitch of a motor before it jams.

00:15:03: Yes, if the AI builder ignores that cassette knowledge and just relies on structured system data—the PLCs and the SCADA —they will build an impressive digital demo that will never survive actual physical production.

00:15:16: AI only scales when it aligns with how factories actually work not how software developers wish they worked.

00:15:22: You can't just bolt an AI onto a legacy process, ignore the operator's instincts and expect magic.

00:15:28: And that collision between advanced software in messy physical reality perfectly sets up our final theme – Physical AI & Humanoid Reality.

00:15:36: The Final Frontier

00:15:37: Because once you have data foundation, simulation twin or operational AI making decisions….

00:15:43: …the last step is embedding all of this intelligence into machines that physically move through facility.

00:15:49: We are moving from Software That Thinks to Hardware That Acts And the momentum here is building incredibly fast.

00:15:55: Daniel Kupfer shared some really eye-opening numbers.

00:15:57: Yeah, China is currently going all in on humanoid robotics backed by roughly twenty six billion dollars and public funding

00:16:04: Twenty six billion.

00:16:06: he highlighted companies like unitary which are seeing massive sales volumes and our targeting multi billion dollar valuations.

00:16:13: But kipper

00:16:14: added a massive caveat to those numbers.

00:16:16: He noted that right now these humanoids are almost entirely being used in research education, and entertainment.

00:16:23: I mean they're doing backflips on video but they are not seeing real scaled industrial deployment on actual factory

00:16:30: floors.".

00:16:30: And Tullish-Renka's post explained exactly why they aren't scaling industrially.

00:16:34: yet it is NOT a technology gap right.

00:16:36: the physical building blocks the actuators The AI models...the spatial computing family They all already here..The failure to scale Is purely an execution problem.

00:16:47: Cherwenka argued that companies are treating physical AI as a simple technology deployment rather than system level execution problem.

00:16:55: You don't unlock value just by purchasing humanoid robot and dropping it onto the floor.

00:16:59: Exactly, you only unlock value when IT systems, operational technologies environments or physical operations are integrated into single-system tied to real performance KPIs.

00:17:10: Consider what Alex Greenberg at Siemens pointed out Pulling from Insights by Michael Walker & Yen Guan The physical reality of a factory is highly unpredictable.

00:17:20: It's chaotic!

00:17:21: A traditional industrial robot sits inside a steel cage and does the exact same weld a million times, but a humanoid has to navigate dynamic environment.

00:17:32: It has to adapt to unpredictable scenarios and interface with legacy machinery that was built thirty years before AI even existed.

00:17:39: Which is why simulation is so critical?

00:17:41: Exactly!

00:17:42: Imagine taking a highly intelligent but totally untrained employee, just dropping them into chaotic thirty year old factory... Four-cliffs are driving by unpredictably, raw materials are stacked unevenly and the legacy machines break down constantly.

00:17:57: Without the digital twin and simulation tools to rigorously test how that humanoid will react to a thousand different unpredictable safety scenarios it is a recipe for an industrial disaster.

00:18:07: Manufacturers have to virtually model these complexities across robotics, safety parameters ,and facility layout before they ever attempt physical deployment.

00:18:18: If you cannot simulate the humanoid's behavior in your specific messy environment, You really have no business putting it on the shop floor.

00:18:25: So when we stack all of these insights together from fixing the basic data plumbing so that systems can speak the same language to simulating entire supply chains and a twin From agentic AI predicting failures To humanoids learning in virtual sandbox Where does this actually leave manufacturing professional today?

00:18:44: I think Brian Marsh offered a perfect synthesis of this landscape.

00:18:47: He issued a very stark warning, ignoring industry.

00:18:50: four point zero is no longer just missing out on it competitive differentiator its

00:18:54: survival.

00:18:55: It's as direct path to structural disadvantage.

00:18:57: companies that resist This connected reality are locking themselves into future rising costs severe productivity losses and supply chain vulnerability.

00:19:06: And beyond the tech they're going face massive talent shortages

00:19:09: Absolutely.

00:19:11: Modern technical talent simply does not want to work in an outdated, siloed analog factory where they have to wait three weeks for a spreadsheet.

00:19:19: tell them what went wrong In twenty-twenty six building is scalable.

00:19:24: data driven architecture is purely about survival.

00:19:27: It really is adapt or die.

00:19:29: but it leaves me with one final thought I want you to mull over.

00:19:32: As factories become these software-defined systems that can orchestrate and adjust themselves in real time, what happens to the human gut feeling that has run manufacturing floors for centuries?

00:19:43: That's a big question.

00:19:44: Does that deep tribal intuition of the operator knowing machine sounds wrong just by walking past it get successfully coded into the agentic AI?

00:19:52: or are we engineering future where human instinct on shop floor becomes entirely obsolete?

00:19:57: It's the ultimate question of whether we are augmenting human capability or replacing

00:20:07: it altogether.