Best of LinkedIn: Robotics CW 32/ 33

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

00:00:07: Frenness 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 the description.

00:00:22: So picture a state of the art multi million dollar robot and it is powered by literally the smartest foundation models available today.

00:00:32: You ask it to unscrew a light bulb and it executes that perfectly like almost every single time.

00:00:36: Which is what you'd expect for the price tag?

00:00:38: Exactly, but then you asked that exact same robot to sweep up some debris with a dustpan And just suddenly fails like seventy percent of the time.

00:00:46: So today we are doing a deep dive into the top robotics trends across LinkedIn To unpack why that happens.

00:00:52: We really want to look at how physical AI Is finally moving out of the you know That polished demo phase and colliding With the actual messy funded reality through infrastructure.

00:01:03: Right, and I think our mission for this deep dive is really to map out how the fundamental constraints on the factory floor have shifted because you know the industry's just no longer impressed by a robot that simply functions in a sterile lab.

00:01:14: Yeah we're past

00:01:15: that We are.

00:01:16: The entire conversation has pivoted to deployment financing And supporting these fleets at scale In really unpredictable manufacturing environments.

00:01:25: You can't talk about physical deployment without first tearing down the quote-unquote brainpowering these machines.

00:01:33: The old debate over whether physical AI is a real category, I mean that's over.

00:01:38: Oh it's definitely over!

00:01:40: It's here right?

00:01:40: Right...the very definition of what constitutes an industrial robot is expanding so rapidly.

00:01:45: we're moving away from just blind repeating hardware.

00:01:49: Yeah, and Andrew Latch pointed to a framework that captures this shift.

00:01:52: really well he noted the Gartner now defines physical AI through three continuous linked capabilities which are perceive reason an act.

00:02:01: Perceive Reason Act?

00:02:02: Okay

00:02:03: right but rather than just viewing this as some you know abstract concept they're actually naming integrators like Accenture and Schneider Electric as market shapers because getting It delivers massive operational games.

00:02:19: How

00:02:19: massive are we talking?

00:02:21: We're talking about twenty five to thirty-five percent faster execution cycles, and that's according to data shared by Nicole Vandette.

00:02:26: Wow a Thirty percent bump in execution speed.

00:02:30: I mean That is the kind of metric that forces manufacturers hand.

00:02:33: Oh absolutely if your competitor has that you don't just you lose.

00:02:38: And the foundation models driving that speed are evolving really aggressively right now, like Google DeepMind just pushed out three new models under their Gemini Robotics umbrella...

00:02:47: Right?

00:02:47: The Gemini releases!

00:02:48: Yeah so there's Gemini robotics two for whole body control ER-II which focuses on environmental reasoning and on device to handle like local low latency processing.

00:02:59: And the emphasis on whole body control is the real lead for there because historically a robotic arm was programmed basically in isolation, right?

00:03:08: Yeah

00:03:08: just the arm doing its one task

00:03:10: exactly yeah.

00:03:11: but now the system coordinates balancing crouching spatial awareness and fine manipulation all through a single learned policy.

00:03:21: It doesn't have to calculate every single joint angle manual anymore, it just intuitively understands its own kinematics.

00:03:28: I have to push back a little bit here though.

00:03:30: let's unpack this gap between the software and physical reality because cinematic launch videos for these systems they look incredible.

00:03:40: but Harshitha Paramishwara Rukmini highlighted and kind of sobering data point regarding that Gemini-powered robot.

00:03:48: It successfully unscrews the light bulb, ninety two percent at a time but the dustpan task it drops to thirty two percent success

00:03:56: rate.

00:03:56: Yeah!

00:03:56: The dust pan problem

00:03:57: Why does this happen?

00:03:58: Its because the lightbulb has rigid predictable geometry.

00:04:01: Once the gripper makes contact rotation is just math.

00:04:04: But a dustpan involves sweeping deformable unpredictable debris friction changes scatter changes

00:04:11: completely different physical interaction.

00:04:13: Right, it's like watching a toddler who can solve this really complex spatial puzzle but then turns around and walks straight into the glass door.

00:04:20: I mean are we getting blinded by polished demos?

00:04:22: We're definitely at risk of that.

00:04:24: Yeah And Tim Iskac actually raised a major warning flag about this specific trap which he calls start-up reality.

00:04:31: Start

00:04:32: up reality

00:04:32: yeah because were flooded with these glossy launch videos perfectly lit pristine corridors.

00:04:41: And it blurs the line between a rendered conceptual vision and the actual grime of daily operations, The phrase we intend to is quietly wearing the costume of.

00:04:51: We already do

00:04:53: that as very dangerous costume when enterprise budgets are on the line?

00:04:56: The tense is just incredibly misleading.

00:04:58: exactly which Is why Kalamaz argues That industrial AI needs To completely abandon the highlight reel.

00:05:04: A factory operates On repeatability not one lucky take for a camera.

00:05:09: He points out that the true evaluation for physical AI has to be scenario-based.

00:05:14: It has lean heavily on detailed telemetry and failure transparency.

00:05:18: A manufacturer needs to know exactly how, why a system dropped a payload—the specific edge case that confused its reasoning model—and how it recovers?

00:05:27: Because reliability is literally the only metric that matters in production lines!

00:05:32: If brain hesitates when line stops you are burning money...

00:05:36: And over fist yeah.

00:05:37: So since the foundation models are clearly still battling with that dustpan level of physical variability, it forces a critical choice for manufacturers today.

00:05:46: How do we deploy them into the real world right now?

00:05:49: Right And I think this brings us to our next theme which is form factor and deciding who actually has to adapt to the mess.

00:05:55: Yeah!

00:05:56: This is absolute crux of the variability problem.

00:05:58: Let's start what bodies were choosing?

00:06:00: Christophe Hamill shared a very clear perspective on form factor selection, and he argued that the industry shouldn't just default to a bipedal humanoid shape.

00:06:09: Just because it looks futuristic

00:06:11: right?

00:06:11: The sci-fi appeal

00:06:12: exactly.

00:06:13: his hierarchy suggests That a bipede robot is really only economically justified when mobility manipulation And then need to navigate human centric environments like stairs and narrow walkways When all of those are required simultaneously.

00:06:27: and comes down to allocating compute power in battery life efficiently.

00:06:32: Keeping a heavy machine balanced on two legs requires constant high-frequency micro adjustments, It is just a massive energy drain.

00:06:42: Why

00:06:42: use the Swiss Army knife when a simple scalpel does the job better?

00:06:45: I mean, for high-speed repetitive work A fixed arm with dexterous tactile hand makes far more sense.

00:06:51: You get human like manipulation without unnecessary complexity of constantly fighting gravity

00:06:57: And that leads directly into broader concept of variability which Paul Vichari brought in to focus When deploying robot facility faces this constant trade off regarding environment itself.

00:07:07: So say you have a vision system that's failing because the shadow falls across a widget at three p.m every single day.

00:07:13: Okay,

00:07:13: pretty common scenario

00:07:15: Right?

00:07:15: Do spend hundreds of hours collecting data labeling it and training an incredibly expensive machine learning model to ignore The Shadow?

00:07:23: Or do just spent twenty dollars To move the overhead light?

00:07:26: I mean moving the lights seems like obvious engineering choice

00:07:29: It is But illustrates fundamental deployment question Does the Machine absorb the variability or does the environment?

00:07:37: Because sometimes, the environment is just too chaotic to bolt down.

00:07:41: Gabriel Millian highlighted Subami Industries' Arcax robot as kind of an extreme example...

00:07:47: Wait!

00:07:47: Arcaxx?!

00:07:47: Is that a massive like mechanical mech suit thing?

00:07:51: The very same.

00:07:51: it's a fourteen point eight foot three and half ton machine with twenty six joints costs around two point seven million dollars.

00:07:58: Wow

00:07:59: But the fascinating part isn't even this size.

00:08:02: It's that there is a cockpit hidden inside, it has piloted robots...

00:08:06: Someone actually sitting in it?

00:08:07: Yes!

00:08:08: Because in unstructured environments like disaster zones or deep mud construction sites conditions change so unpredictably.

00:08:17: The most practical solution right now isn't removing human and waiting for AI to catch up its designing heavily armored seats so they can leverage the machine's sheer strength.

00:08:27: That is a serious reality check for anyone assuming full autonomy.

00:08:30: as just like, A software update away?

00:08:32: Yeah!

00:08:33: But to be fairer on the software side there are real world examples where brain is successfully absorbing that variability.

00:08:39: today Ralph Gould shared that Sarriact is running a single robotic control system for Arvato across three totally different fulfillment sites.

00:08:48: Dortmund, Gooterslow and Memphis.

00:08:50: Three totally different layouts

00:08:52: Right!

00:08:52: And they didn't have to rebuild the programming of each specific warehouse.

00:08:56: The Single System reads the unstructured mess in front it and adjusts its routing on the fly

00:09:01: which represents basically the holy grail of entrologistics.

00:09:04: When a system can adapt to a stray pallet or changing light condition, you know moving forklift without an engineer stepping into right new rules.

00:09:14: it transitions from pilot project to true industrial standard.

00:09:18: Speaking transitioning.

00:09:19: if your finding this breakdown of robotics landscape valuable take second hit subscribe.

00:09:25: so don't miss future editions.

00:09:29: So we've established that fixed arms and pragmatic single-brain logistics systems sort of dominate practical deployments today.

00:09:37: Yet if you look at the capital markets, the funding is aggressively pointing in a completely different direction.

00:09:42: The money is overwhelmingly backing humanoids as the ultimate long term hardware platform.

00:09:48: Yuri Rebrek noted an incredible statistic Eighteen point eight billion dollars flowed into robotics in just the first half of twenty-twenty

00:09:57: six.

00:09:57: That's Just

00:09:57: The First Half, and fifteen startups In This Space now hold a collective valuation Of Over One Hundred Billion Dollars.

00:10:05: And this standout example there is obviously figure AI Which Is Carrying A Thirty Nine Billion Dollar Valuation.

00:10:11: Xeonam Provided An Excellent Breakdown On How The Market Is Actually Justifying that Number Because Investors Aren't Valuing Figure As A Traditional you know, metal and motors hardware manufacturer.

00:10:22: They are pricing it as a full stack embodied AI platform

00:10:26: so they're betting on the operating system essentially

00:10:28: specifically their Helix system.

00:10:30: yeah The goal of helix is to scale complex physical behaviors using natural language prompts rather than manual line-by-line programming.

00:10:38: .The market is pricing in a generalist model.

00:10:40: if It works on the BMW factory floor today ,the option value suggests that easily transitions into logistics tomorrow And eventually into consumer homes.

00:10:49: But let's pause and look at the physics of that ambition for a second.

00:10:52: Everyone is sewing billions at foundation models in AI brains, but what happens when these incredibly complex highly articulated machines actually have to run a grueling eight hour shift lifting heavy automotive parts?

00:11:04: Yeah then you hit.

00:11:08: Dale Tutt addressed this fundamental constraint.

00:11:10: The ultimate barrier to humanoid adoption might have absolutely nothing do with the intelligence of AI, it may simply be batteries.

00:11:18: The thermal management alone for a continuous bipedal movement must an absolute nightmare!

00:11:23: It is a profound engineering challenge.

00:11:26: as the AI models become more complex their local computing requirements spike which obviously drains power.

00:11:32: Add in the energy required to fire dozens of actuators just to maintain a standing balance, let alone lift a payload and the power draw is immense.

00:11:40: TUNT emphasizes that runtime charging speed and energy density.

00:11:44: these aren't secondary features.

00:11:45: they are the baseline requirements for untethered deployment.

00:11:48: if general purpose humanoid has to remain tethered into wall outlet for power it entirely loses mobility.

00:11:55: that justifies its form factor.

00:11:57: It always comes back to the constraints of physical hardware.

00:12:01: And this hardware bottleneck is also shaping global ecosystem race we're watching unfold right now.

00:12:07: Clary Chao mapped out the geopolitical division of labor really clearly, The US is driving AI foundation models and edge computing architecture.

00:12:15: China is leveraging its massive manufacturing scale To create unparalleled hardware data flywheel And Europe is acting as the anchor for industrial automation and complex factory integration.

00:12:27: The winner won't be a single company building a standalone robot, it'll be whichever region successfully orchestrates all three of those pillars...

00:12:34: ...and the sheer scale China's manufacturing push is just hard to overstate.

00:12:39: Aaron Prather noted massive strategic bet by Mitsubishi Motors.

00:12:42: They are actually taking an unused engine production line at their cuto plant entirely repurposing it mass-produced AI powered humanoid robots

00:12:50: in GenLine

00:12:52: Starting in early twenty-twenty seven, their target is one thousand humanoids every single month and they are deploying them inside their own facilities first to combat severe labor shortages.

00:13:01: I mean pumping out a thousand units a month

00:13:03: completely

00:13:04: upends the traditional economics of robotics.

00:13:07: it drives down the preunit cost of actuators and sensors To point where humanoid shift from these capital intensive luxury items to standard operational expenses.

00:13:17: So we have brains getting smarter.

00:13:18: We have bodies scaling into mass production billions in capital kind of lubricating the gears, but that leaves the final and perhaps most difficult hurdle which is training.

00:13:29: Yes!

00:13:30: How does a legacy manufacturer teach a fleet of five hundred new robots without hiring a permanent incredible expensive army of software engineers to just live on the shop floor?

00:13:41: That question cuts straight to the core.

00:13:44: Dave Saunders highlighted two startups that launched practically simultaneously, but with entirely opposite approaches to this exact training bottleneck.

00:13:52: On one side you have Walden Robotics which is backed by Toyota and has three hundred million dollars in funding.

00:13:58: their model involves building the full application stack internally.

00:14:01: so they're own engineers are The ones who teach the robot every new behavior required by the client.

00:14:06: So they aren't just selling hardware.

00:14:08: They are essentially a professional services firm With a robot attached to the invoice.

00:14:13: That is exactly the implication.

00:14:15: Their scaling limitation is directly tied to human headcount, but on the other side you have reimagined robotics founded by a former Google DeepMind lead.

00:14:25: their strategy is entirely different.

00:14:27: they empower The Human Operator-the person already doing the manual job On the factory floor To teach the robot through physical demonstration

00:14:35: Like guiding the arm to show it the path rather than coding the coordinates on a laptop?

00:14:40: Precisely They managed to cut the prototyping time for a complex electronics disassembly task from a full day of engineering down to about ten minutes.

00:14:48: Wow!

00:14:49: That fundamental difference dictates whether you are selling service contract or scalable tool.

00:14:55: If an operator can retrain machine in ten minutes between shifts, manufacturer owns capability.

00:15:00: if engineer has fly out rewrite script manufacturers dependent forever

00:15:04: And Europe is making a highly strategic, well-funded move to ensure their manufacturers maintain that independence.

00:15:12: And Tutova reported on David Rager's Nirae Robotics which recently secured one point four billion dollars.

00:15:18: Billion with the B?

00:15:19: Yes

00:15:20: it is being called The Largest Full Stack Robotics Funding Round in History.

00:15:24: Their primary initiative with that capital is building what they call Niraea gyms starting from massive facility at RWTH Akin University.

00:15:32: I

00:15:32: find this concept so fascinating.

00:15:34: The premise is that pure AI intelligence is no longer the primary bottleneck, the bottleneck has tacit historical experience.

00:15:41: a legacy factory holds decades of proprietary messy real-world knowledge regarding how materials bend How lighting shifts?

00:15:48: How specific machines vibrate.

00:15:50: No generalized AI model in Silicon Valley possesses that granular physical data.

00:15:54: Exactly So new grade gyms provide infrastructure to capture it.

00:15:58: They use system called brain accounts.

00:16:01: A manufacturing company can utilize the gym's resources to train and fine-tune their own physical AI model based on specific processes.

00:16:09: The critical distinction here is intellectual property, the factory retains full ownership of that trained knowledge.

00:16:16: So they aren't just feeding a tech giant centralized model for free.

00:16:20: Correct, the manufacturer packages their hard-earned operational knowhow into this brain account and can then deploy those learned skills across their entire internal fleet.

00:16:30: It locks the value inside The Industrial Company rather than the software provider.

00:16:35: it is A massive paradigm shift For industrial data strategy.

00:16:39: the integration between the physical hardware And these proprietary software brains Is becoming incredibly tightly coupled.

00:16:45: Leo Su pointed out a major corporate move reflecting this.

00:16:48: Deepseek, which is one of the premier AI foundation model companies in China took a two point three-one percent stake and Unitree's recent IPO.

00:16:56: it was a massively over subscribed offering.

00:16:59: but deep seek isn't just acting as passive capital here.

00:17:02: they are locked in for thirty six months in a joint development pact to build The Brain For Unitrees hardware.

00:17:08: It Is A Direct Structural Fusion Of AI Intelligence And Mechanical Engineering.

00:17:13: But despite all these technological alliances and massive funding rounds, the final barrier on the factory floor is surprisingly mundane.

00:17:22: Moritz Frender points out that the true bottleneck for enterprise adoption isn't a lack of technical ambition anymore.

00:17:29: The technology while still imperfect is viable –the actual friction-is entirely organizational.

00:17:35: In a word yes!

00:17:37: Who holds budget for general purpose humanoid?

00:17:39: Is it IT?

00:17:40: Because its essentially a massive edge computing device.

00:17:43: Or is it operations because its replacing manual labor?

00:17:46: Furthermore, who absorbs the risk if an autonomous two hundred pound machine drops heavy payload on crucial conveyor belt?

00:17:52: Yeah

00:17:53: that's nightmare scenario.

00:17:54: Right and once fleet is live what does operating model look like to maintain patch repair them?

00:18:00: Those structural questions paralyze procurement departments

00:18:03: The ultimate corporate friction.

00:18:06: So what does this all mean for the professionals actually trying to build and manage these facilities?

00:18:11: The robotic space has officially graduated from lab experiments to robots as a service.

00:18:16: The winners in this upcoming cycle will not merely be the companies with the shiniest titanium hardware, but also the integrators who can seamlessly plug into existing factory software, engineers that crack battery and thermal constraints for true untethered eight-hour shift or most importantly platforms that empower floor operators train their own machines.

00:18:36: I

00:18:38: will leave you with a final thought to mull over as this ecosystem matures.

00:18:42: We are moving toward the future where robotics hardware becomes increasingly commoditized, and platforms like NierA Gems allow manufacturers build and license their own brain accounts!

00:18:52: If that happens does the ultimate competitive advantage shift away from high-tech robotic startups in Silicon Valley?

00:18:59: Does the real power move over to legacy manufacturers simply because they hold decades of messy, invaluable proprietary data

00:19:17: on how?