Best of LinkedIn: Robotics CW 30/ 31

00:00:00: Provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn post about robotics in calendar weeks thirty-and-thirty one.

00:00:07: Frenness is a B to B 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.

00:00:21: You know imagine spending millions of dollars to integrate like the most advanced AI brain in the

00:00:26: world

00:00:27: into A new robotic rig.

00:00:29: Right.

00:00:30: A massive investment

00:00:31: Exactly, and then you ask it to perform a really basic factory task Something as simple as marking a clapperboard

00:00:38: Which should be easy for a multi-million dollar system.

00:00:40: But instead of just gently completing the task The robot violently snaps its physical testing arms backward.

00:00:47: Oh wow Yeah

00:00:48: It completely destroyed the hardware.

00:00:50: And I mean that isn't some hypothetical scenario.

00:00:52: That actually happened.

00:00:53: We're gonna break down exactly why today.

00:00:55: Yeah, I mean if you're managing a production line or dealing with industrial automation today.

00:00:59: You know just how incredibly noisy the robotics space has become.

00:01:04: Oh it's just a constant flood of press releases.

00:01:05: yeah completely.

00:01:07: so our mission With this deep dive is to Cut right through that noise for you.

00:01:13: We've curated the most critical robotics trends straight from The front lines on LinkedIn over the past couple of weeks.

00:01:20: Yeah, no fluff just the raw intelligence You need for the factory floor

00:01:23: exactly.

00:01:24: and I guess to understand the physical robots hitting the floor we really have To start by talking about the software driving them.

00:01:31: yeah

00:01:32: because?

00:01:34: I mean, traditional automation was entirely bespoke.

00:01:40: Like writing custom code for one specific arm in one specific spot?

00:01:44: Exactly!

00:01:45: But now as Lucas M. Ziegler and Timmis Kack highlighted we're looking at things like Google's Gemai Robotics too.

00:01:53: right they are writing new code for every machine anymore.

00:01:57: They're actually taking a single model checkpoint and running it across entirely different robot

00:02:01: bodies.".

00:02:01: Okay, wait let me just stop you there because that term uh single-model check point...it gets thrown around a lot?

00:02:07: It really

00:02:07: does!

00:02:08: What does that actually mean in the physical context?

00:02:11: like if an AI is trained on internet text or static images how does that snapshot of data suddenly know how to?

00:02:18: I don't know.

00:02:18: balance onto legs?

00:02:19: yeah That is literally the core friction point of the entire industry right now.

00:02:25: So a model checkpoint, it's basically a saved snapshot of an neural network brain after massive amounts training and theory behind this whole body.

00:02:36: reasoning that if AI understands the general physics logic you can plug single-brain into any hardware.

00:02:45: It

00:02:46: will figure out how to control the system fluidly rather than you having to write rigid, separate programs for like the left arm and right arm in legs.

00:02:55: I mean that sounds highly efficient on paper?

00:02:58: As a plant manager i honestly don't care about what happens when you flip this switch which brings us back into that broken testing rig at start.

00:03:06: Oh

00:03:06: yeah!

00:03:06: The clapperboard incident.

00:03:07: Right.

00:03:08: EraZoo shared an evaluation recently just perfectly illustrates this gap between AI logic and physical reality.

00:03:16: They actually pitted Anthropix Claude Opus V against Google's Gemini ER-II on real robotic

00:03:23: tasks.

00:03:24: Yeah, and the results were pretty shocking.

00:03:26: Claude scored a one out of five on that clapperboard test...

00:03:29: Which isn't great but it managed to do

00:03:31: it?

00:03:31: Right!

00:03:32: It struggled but completed it.

00:03:33: But Gemini scored is zero out of Five And like we said didn't just fail.

00:03:38: it literally output a joint command That forced the YAM testing arms to move in a way that hardware physically couldn't support.

00:03:46: Just

00:03:46: completely flipping the rig over and breaking it?

00:03:49: Exactly!

00:03:49: Which, honestly...it kind of blows my mind.

00:03:51: we have these models can write flawless python code in two seconds but they don't know.

00:03:56: pushing metal arm past joint limit will snap

00:03:59: it.

00:03:59: It's wild

00:04:00: right?!

00:04:01: Why wouldn't developers just, you know hard-code a basic safety limit?

00:04:04: Well

00:04:04: because the second you start hard coding those rigid safety boundaries You kind of lose The fluid general purpose adaptability that you wanted from the AI in the first place.

00:04:13: Oh right.

00:04:14: Because your back to bespoke coding.

00:04:16: Exactly, the issue is that large language models they process language and logic.

00:04:21: They don't have this inherent mathematical understanding of things like torque or friction Or gravity.

00:04:27: right

00:04:27: if chat GPT gives you a wrong answer It's just a bad string of text.

00:04:30: yeah

00:04:31: But in the physical world outputting the wrong spatial coordinate Breaks a fifty thousand dollar piece of hardware.

00:04:37: So it's basically an extrapolation problem.

00:04:40: The AI knows what the end goal looks like but It fundamentally lacks the intuitive physics to calculate the force needed To get there safely perfectly

00:04:49: said

00:04:49: and I imagine that just gets infinitely worse when you factor in reaction time, right?

00:04:53: Like a human doesn't calculate gravity When catching a falling tool we just grab

00:04:57: it.

00:04:57: yeah reaction Time is a massive structural bottleneck right now.

00:05:01: Yeah And shamas Khan actually brought up a new framework from MIT and other universities called VLAH that tries to solve this.

00:05:09: Oh, VLA H!

00:05:10: How does that work?

00:05:11: So...to get why it's a breakthrough you have to look at normal vision language action models.

00:05:17: they run on the synchronous loop The robot's camera takes a picture sends it to the brain the brain calculates the move And then send the command of the motors.

00:05:25: But all that computing takes what a fraction of a second exactly.

00:05:29: And in that fraction of a second, the physical world has kept moving.

00:05:32: Right!

00:05:32: So by the time the motor gets to command... ...the part robot was reaching for isn't even there anymore?

00:05:37: Yep

00:05:37: so VLAH changes the math entirely.

00:05:40: Instead of reacting at present moment The model actually calculates its own computational delay.

00:05:46: Oh wow

00:05:47: Yeah.

00:05:47: and it predicts the robots future execution state.

00:05:50: Okay, wait so it's like a quarterback throwing a football.

00:05:53: you don't throw the ball to where the receiver is right now You calculate their speed your arm strength and you throw to the empty space Where they're gonna be in two seconds.

00:06:01: That Is A perfect analogy.

00:06:03: VLAH Plans The Motor Actions For Where The Robots Joints Will Physically Be by the time that command actually hits the hardware.

00:06:11: That's brilliant!

00:06:12: It really is.

00:06:13: By structuring the math, it cuts reaction latency up to eleven point eight times.

00:06:19: We're suddenly seeing those videos of robots playing ping-pong against humans.

00:06:24: Okay so predicting future to fix latency is huge but let's zoom out from specific joint movements to navigating an actual factory floor

00:06:33: right at the macro level

00:06:34: because Even with perfect reaction time, a robot is totally useless if it gets lost in... ...a five hundred thousand square foot warehouse.

00:06:42: Oh completely!

00:06:43: And Kalmos shared this fascinating finding about spatial navigation.

00:06:47: that really flipped my perspective on how these models see.

00:06:51: Researchers fed models these incredibly dense detailed three-D depth maps to help them navigate which

00:06:57: you'd think would be perfect.

00:06:58: Right, but it barely improved performance at all.

00:07:00: Yeah because processing a massive three-D point cloud of factory floor just eats up an immense amount of compute.

00:07:06: the AI basically gets overwhelmed trying to parse every single pixel of depth.

00:07:10: It's too much data.

00:07:11: But when they stripped that away And this is the crazy part and just gave them model text-based directional hints like typing out the compass degrees of its current heading.

00:07:21: Yeah, performance absolutely skyrocketed on a grasping task.

00:07:24: The success rate jumped from six percent to thirty two percent Just by telling me AI which way it was facing.

00:07:30: It's amazing if really reveals a flaw in how we assume these models operate.

00:07:36: We just assumed they need hyper realistic digital vision right?

00:07:40: Like human eyes

00:07:41: yeah but actually a simple vector like a compass heading, it grounds the model's spatial awareness instantly.

00:07:48: Knowing your exact orientation in a coordinate system is vastly more useful to the machine than having high-def, three D picture of a forklift driving by.

00:07:56: It turns out a compass beats a fancy depth map!

00:07:59: Knowing where you are matters more then having perfect pictures of it...

00:08:02: Exactly

00:08:02: Which naturally brings us into what I think as ultimate reality check this whole industry which is

00:08:08: The Sim to Real Gap.

00:08:09: Yeah

00:08:09: Jerry Huang who runs robotic center in Sanford made point that every manufacturing pro needs hear.

00:08:17: The bottleneck right now isn't that the models aren't smart enough.

00:08:20: It's that the training data is fundamentally flawed because it relies way too much on Simulation and teleoperation,

00:08:28: right?

00:08:28: Because we have these incredible physics engines.

00:08:30: Now like Isaac sim you can put a digital robot in a digital factory And run millions of iterations.

00:08:37: But a digital simulator with completely sanitized reality totally.

00:08:41: if I have a real robotic arm running In a real Factory after four hours of lifting The motors heat up, you get thermal throttling.

00:08:48: The physical gears experience backlash

00:08:50: A camera sensor drifts by a millimeter.

00:08:52: Yeah Exactly!

00:08:53: A simulator just doesn't capture the messy degrading physics of actual hardware in dirty environments

00:08:59: And people think that can fix it By having human operator put on VR suit and physically demonstrate to robot.

00:09:07: Right, tele-operation

00:09:09: But generates what we call off policy data.

00:09:11: Okay, explain off policy for a second because I know a lot of people see those videos Of humans driving robots and they assume the training problem is solved

00:09:19: right.

00:09:20: well When a human controls a robot via VR, the robot isn't actually learning its own physical limits.

00:09:26: If the robot starts to tip over... ...the human operator instinctively shifts their own weight in the suit….

00:09:32: …to correct it!

00:09:33: Ah so the balance correction comes from humans' inner ear basically?

00:09:36: Exactly – The Human Vestibular System not the Robot's sensors.

00:09:40: So data is off policy.

00:09:42: It represents a human controlling machine, not the machine learning to autonomously control its own mass.

00:09:49: So the only way actually close that gap is through real world failures?

00:09:52: Basically yeah But real-world failures cost real money.

00:09:55: I mean Gabriel Milliam posted video recently of a unitary robot navigating some really complex terrain and To learn that maneuver The robot fell ten thousand times in a digital simulation.

00:10:05: Ten thousand time Yeah.

00:10:07: And why?

00:10:08: because dropping pixels off a digital rock is totally free.

00:10:12: But if I'm running a logistics hub and i deploy that robot on a real loading dock in February with ice everywhere, and forklifts rushing by... Oh man yeah.

00:10:20: If it falls there- That's not a learning opportunity!

00:10:23: That is a massive OSHA violation.

00:10:25: It destroys cargo..it halts the supply

00:10:27: chain.

00:10:27: Right.

00:10:28: so a demo trained purely in simulation just proves of very narrow capability.

00:10:33: It proves absolutely nothing about readiness For a chaotic messy environment.

00:10:38: Which leads us to The biggest question If the software is still snapping test arms and struggling to get out of the simulator, how are we seeing constant press releases about humanoids actively working in factories today?

00:10:49: Yes.

00:10:49: For fair question let's look at actual deployment scorecards right now.

00:10:53: Let's do it!

00:10:54: Unimatured shared some hard numbers on figures' humanoid pilot program over at BMW.

00:10:59: This wasn't a lab-test.

00:11:01: One of their robots logged twelve hundred and fifty operating hours, kept an accuracy rate above ninety nine percent.

00:11:07: And successfully helped build thirty thousand cars!

00:11:10: I mean those are the kinds numbers that look fantastic on a slide deck...

00:11:13: But let's apply some plant manager math to that okay?

00:11:16: An auto-plant runs two hundred forty seven.

00:11:19: there are one hundred sixty eight hours in week.

00:11:21: So, twelve thousand two hundred fifty hours is basically a month and half of continuous operation for one workstation.

00:11:28: It's barely a blip!

00:11:30: Right it's an incredible engineering milestone sure but...is the six week pilot material success or just very expensive highly supervised proof-of concept?

00:11:40: Exactly

00:11:41: And contrast that with Tesla's Optimus program Elon Musk actually admitted on recent earnings call that Optimus still not doing what he called Material Factory Work

00:11:50: which Pretty revealing.

00:11:51: Yeah, they're currently deployed mostly for data collection and learning.

00:11:54: They aren't carrying the burden of real production quotas yet.

00:11:56: But

00:11:57: then again we are seeing different approaches globally.

00:11:59: Ken Kwong and owner says Gin noted that Xiaomi's humanoids Are actively handling parts in transporting materials inside their EV factories right now.

00:12:07: So there is this massive divergence.

00:12:08: Some were doing highly controlled short-term pilots while others are supposedly integrating them directly into actual material transport.

00:12:15: Yeah.

00:12:16: And this raises a fundamental question about the economic model

00:12:26: of

00:12:45: every single shift.

00:12:46: Right, and her argument is that a general purpose humanoid one built to do anything Is basically build-to-do nothing perfectly in an industrial context.

00:12:56: It's a jack of all trades master of none.

00:12:59: Exactly if you take a hyper expensive General Purpose Humanoid And just drop it into a workstation That was ergonomically designed for a human

00:13:07: Without rethinking the workflow.

00:13:08: Yeah, without fundamentally Rethinking the assembly line you're likely just bleeding capital

00:13:12: totally.

00:13:13: and by The way for the manufacturing pros listening if You want to keep tracking which of these deployments actually stick?

00:13:19: And turn into long-term ROI make sure To subscribe so you catch our future deep dives.

00:13:23: yeah We're definitely keeping a close eye on.

00:13:25: they transition from pilot to real production For

00:13:28: sure and part about transition involves really questioning the hardware design itself.

00:13:33: We just assume these robots need to look exactly like us,

00:13:36: right?

00:13:36: The whole humanoid shape.

00:13:37: but do industrial robots actually need five-fingered hands?

00:13:41: Michelle Sun highlighted tactic systems.

00:13:44: they're building robotic hand specifically for manufacturing dexterity Like electronics assembly But they intentionally only gave the hand three fingers

00:13:53: because it's vastly more practical From a maintenance perspective, A three-finger hand has fewer complex actuators.

00:14:00: Yeah.

00:14:01: Fewer potential points of failure on the line less downtime dramatically lower replacement costs.

00:14:07: You just don't need a biological complexity of human hands to pick up a spark plug

00:14:10: Exactly!

00:14:11: But here's critical insight and I loved this.

00:14:14: Having most practical hardware specs whether it is three fingers or perfect cost profile that doesn't guarantee factory will actually adopt.

00:14:22: Oh not at all.

00:14:23: Dave Saunders shared a parallel from the surgical robotics sector that just perfectly illustrates this hidden friction.

00:14:29: The Johnson & Johnson Surgical Robot, right?

00:14:31: Yeah.

00:14:32: J&J's Atovo Robot—just one FDA authorization and by almost every technical metric on paper it is a superior machine to the current market incumbent which is Intuitive Surgicals DaVinci System.

00:14:44: Okay

00:14:45: so its better!

00:14:46: It's smaller...it has better automation features....and it's backed by J&K.

00:14:50: Yet, market analysts largely agree it will not put a dent in Intuitive's Market Share anytime soon.

00:14:55: Wait!

00:14:56: If the Ottawa is objectively The Better Machine why wouldn't hospital admins just

00:15:00: switch?

00:15:02: Because Intuitives' primary defense isn't their spec sheet – Their Defense Is Surgeon Habit.

00:15:07: There are tens of thousands highly specialized surgeons who have decade muscle memory physically wired to Da Vinci Consul.

00:15:16: Switching systems isn't just a financial purchasing decision.

00:15:19: It requires asking a highly skilled professional to completely unlearn, a complex life or death task that their hands currently perform on autopilot.

00:15:27: That is the unlearning mode Yeah.

00:15:29: And that exact same logic applies to the factory floor.

00:15:33: If I buy a fleet of state-of-the art humanoids...I'm not paying the invoice for hardware.

00:15:38: I am paying the massive operational cost of telling my safety inspectors, my floor managers and unionized line workers to completely abandon a decade of established protocols.

00:15:49: You have retrain the entire human ecosystem around the robot?

00:15:53: The friction isn't hardware integration it's the human behavioral change.

00:15:57: Exactly!

00:15:57: You either design a robot that perfectly mimics existing workflow so switching is invisible or you just accept deployment will be incredibly slow

00:16:07: with that level of human friction, plus software that still struggles with real-world physics.

00:16:14: You really have to ask who is actually footing the build.

00:16:16: solve this?

00:16:17: Right Because the capital requirements are just astronomical.

00:16:20: Oh it's a trillion dollar arms race.

00:16:22: Nicola Lazaroff and Titus Golbus highlighted that global robotics funding shit.

00:16:26: eighteen point eight billion dollars in just six month period.

00:16:29: Eighteen

00:16:30: point eight million.

00:16:31: Yeah!

00:16:31: That more capital injected half year than we've seen previous full years.

00:16:35: And we're seeing an aggressive push in Europe too.

00:16:37: I mean, Nurea Robotics closed a one point four billion dollar series C giving them seven billion valuation.

00:16:44: Incredible!

00:16:44: Humanoid, the European startup just raised a hundred and fifty two million dillian Series A basically positioning them to be Europe's first humanoid unicorn.

00:16:54: Meanwhile In The US Agility Robotics is going public via a SPAC merger

00:16:59: Making Them The First publicly listed, dedicated humanoid company.

00:17:04: Right which is going to force a massive amount of transparency onto the actual unit economics of building these things.

00:17:09: But wait why are European companies like NURA raising at a seven billion dollar valuation when they aren't The ones physically manufacturing and shipping tens of thousands of units yet?

00:17:19: It's wild right.

00:17:20: Marcel Munch provided some crazy contrast here by looking at the Chinese market.

00:17:24: Okay The divergence in value captures is fascinating.

00:17:27: Western firms are increasingly pitching themselves as the intelligence layer, like Newer A they aren't just building physical robots They're building digital infrastructure Like their Newer a gyms where multiple robotic platforms train in simulation

00:17:40: So that trying to build the software mode

00:17:42: Exactly, whereas in China the hardware is becoming a rapidly commoditizing scale game.

00:17:47: Yeah Much pointed out that Agibot which is currently the world's largest shipper of humanoid robots by volume By the way Is targeting an IPO valuation of roughly five point one to six point four billion

00:17:58: So lower than new rise valuation even though Agibott is actually pushing massive physical volume at the door.

00:18:05: Yep

00:18:06: The Chinese supply chain has just brutally cutthroat right now.

00:18:09: You have companies like Agibot, Unitry, UB Tech fighting this vicious price war driving the cost of capable humanoids down to a fraction what Western models costs

00:18:19: which I mean that creates massive geopolitical and economic

00:18:23: tension

00:18:24: because if i'm plant manager looking at my margins Chinese manufacturers are mass producing highly capable humanoid backed by an unstoppable supply chain dropping the unit economics to twenty or thirty thousand dollars.

00:18:38: Why wouldn't Western factories just buy a fleet of those tomorrow and automate their lines?

00:18:42: Well, because the regulatory landscape is rapidly slamming the door on that exact scenario.

00:18:45: Oh

00:18:45: really?!

00:18:46: Yeah Evan Beard noted that the FCC recently banned the import-and sale Of new foreign made advanced robots.

00:18:53: Wait!

00:18:53: Just a blanket ban.

00:18:54: On what grounds?

00:18:55: Purely

00:18:56: from a national security & infrastructure standpoint.

00:18:59: It specifically includes humanoids Quadrupeds and autonomous mobile robots.

00:19:04: Wow!

00:19:05: And Robert Little actually shared a specific report where researchers documented a Chinese Unitry G-One robot, actively transmitting its onboard sensor in system data to servers in China every five minutes.

00:19:17: Every five

00:19:18: minutes?

00:19:18: Yeah completely transparent for the operator.

00:19:20: We are talking about a mobile machine equipped with high definition cameras LIDAR and depth sensors constantly mapping its environment.

00:19:28: Precisely, and the researchers also identified vulnerabilities that could theoretically allow for remote takeover of a machine.

00:19:34: Oh

00:19:35: man

00:19:36: Now obviously it's important to state we are simply reporting findings detailed in these sources but underlying logic is pretty clear.

00:19:43: These aren't just dumb robotic arms bolted into concrete floor anymore.

00:19:47: Right they're highly mobile hyper connected sensor platforms

00:19:50: Exactly.

00:19:51: If you deploy hundreds of these inside a critical manufacturing facility, an aerospace plant or logistics hub and that high-res mapping data is exposed right?

00:20:01: Or if the physical machine itself can be compromised

00:20:04: The physical and economic risk is catastrophic.

00:20:06: Yeah, the US government essentially made the calculation That it just cannot allow mobile platforms It doesn't entirely trust to map its manufacturing base regardless Of how cheap the hardware is.

00:20:17: And, you know it actually goes beyond just geopolitical hardware bands.

00:20:21: It extends to software sovereignty as well.

00:20:23: Oh definitely Birkin Killick shared that a company called Micro-Aji Just signed the open weights letter and their argument perfectly encapsulates this tension.

00:20:33: What do they argue?

00:20:34: Basically if your manufacturer running plant with high robot density those robots are quite literally means of production.

00:20:41: Right!

00:20:42: If don't own actual weight of AI model running machines If you're just renting that intelligence via a cloud API from some massive tech giant...

00:20:52: And your entire physical operation is dependent on third-card server.

00:20:56: Exactly!

00:20:56: Whoever controls the AI model, controls the factory?

00:20:59: That's a sobering thought.

00:21:00: It really is.

00:21:01: So we've journeyed from AI models violently snapping physical testing arms because they don't understand gravity to the massive friction of making human workers unlearn their habits

00:21:12: All the way up to billion dollar funding rounds and absolute Bands on foreign hardware mapping our infrastructure.

00:21:19: Yeah,

00:21:19: the technology is moving at a blistering pace but The physical and human constraints are just Incredibly stubborn which

00:21:26: leaves us with a critical paradigm shift to consider.

00:21:29: I think

00:21:29: exactly.

00:21:30: i want To leave you With a final thought to mull over as You look At your own operations When you combine the lessons of that brutal, sim-to-real gap where digital training shatters against physical reality with the massive unlearning cost we saw in the surgical robotics example.

00:21:45: In this coming wave of physical AI?

00:21:48: will the market winner actually be the smartest most physically capable humanoid?

00:21:53: Or would it simply be a machine that requires absolute least amount behavioral change from humans working next to?

00:21:59: I

00:21:59: mean, that is the multi-billion dollar question every robotics founder needs to answer right

00:22:19: now.