Best of LinkedIn: Robotics CW 44/ 45

00:00:00: Provided by Thomas Allgaier and Franis, based on the most relevant LinkedIn posts about robotics in calendar weeks forty-four and forty five.

00:00:08: Franus 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.

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

00:00:22: Right.

00:00:22: today's deep dive has really tailored specifically for You The Smart Build & Manufacturing Professionals out there

00:00:28: Yeah, the ones actually doing their work.

00:00:29: Exactly!

00:00:30: We are extracting the absolute top robotics trends from our curated LinkedIn sources and you know we're cutting straight through all that promotional fluff.

00:00:39: No marketing spin today

00:00:40: Right I'm going to look at real world deployments The massive underlying data bottleneck How China is just aggressively scaling And why the most profitable automation move right now might be the most boring one.

00:00:54: I love that.

00:00:55: So let's start with the most visible trend, right?

00:00:58: Humanoid robots... Oh yeah They are officially moving out of those carefully curated you know highly edited demo videos.

00:01:05: Finally Right and they're actually stepping onto active factory floors.

00:01:10: Philip Veen shared this really fascinating insight about this shift.

00:01:13: Now what did he note?

00:01:15: so Bosch is taking a massive step by actively manufacturing humanoids at their facility in Buhl.

00:01:22: Yeah

00:01:22: And they are leveraging their one hundred and forty year legacy of building sensors, motors and actuators.

00:01:28: See that is huge!

00:01:29: Because?

00:01:30: They're bringing automotive scale manufacturing to this.

00:01:33: Exactly

00:01:33: Which the exact capability like almost every robotic startup lacks right now.

00:01:37: Right because a start-up can hand build ten amazing prototypes in a lab.

00:01:41: Yeah

00:01:41: In a pristine perfect lab

00:01:42: But completely lack supply chain to build ten thousand reliable units That won't just shake apart after month.

00:01:48: industrial vibration

00:01:49: Exactly.

00:01:50: And, you know we are seeing the software side trying to match that industrial ambition too.

00:01:55: Lukas M. Zickler shared a major update on Aptronix Apollo II humanoids.

00:02:00: All right!

00:02:01: The ones powered by Google DeepMind?

00:02:03: Yeah Gemini Robotics.

00:02:04: two point five they are now actively doing automotive kidding online.

00:02:08: Mercedes Benz lines in both Germany and Hungary...and the headline number their pushing is a ninety six percent success rate.

00:02:16: okay wait I have to push back Do it.

00:02:20: Because we really need to define what automotive kidding actually is physically,

00:02:25: right?

00:02:26: It's not just moving identical boxes.

00:02:27: exactly.

00:02:28: its a robot looking into multiple bins identifying differently shaped parts like A metal bracket or rubber gasket some bolts and arranging them in a tray

00:02:38: Right for the final assembly line.

00:02:39: So at ninety six percent success rate per pick sounds great in a press release But Guillermo floor posted a breakdown of this math that just completely changes the calculus

00:02:49: Really?

00:02:50: How so

00:02:51: well if a robot is picking twenty different parts to finish one kidding tray Okay, and it has four percent failure rate on every single pick?

00:02:59: Yeah The compounding probability means that the completed tray only comes out perfectly like forty-four percent of the time.

00:03:06: Oh wow Right really

00:03:07: yeah.

00:03:08: So if you're plant manager listening You aren't buying in ninety six percent average your looking at the worst case error right

00:03:15: right the bottleneck.

00:03:16: uh, fifty-six percent failure rate creates on like a Friday night shift.

00:03:21: Exactly it's a brutal reality check.

00:03:23: Yeah but I mean doesn't that just mean the ninety six percent isn't the real breakthrough here?

00:03:28: Right and its not.

00:03:30: The Real Breakthrough is the learning speed.

00:03:33: Abertronik noted these models improved dramatically in just few weeks of post training.

00:03:37: Ah okay so It' s the speed at which software learns to close four percent gap

00:03:42: exactly.

00:03:43: But if complex tasks like kidding are too risky for widespread adoption right now, it makes you wonder where early buyers are actually finding value.

00:03:51: Right because these things are massive capital expenditures.

00:03:55: Well Erdem Uren actually highlighted this.

00:03:58: Early European buyers like BMW Renault Eon They aren't doing complex assembly.

00:04:04: What're they doing then?

00:04:05: They starting with the most unglimerous logistics task.

00:04:08: You can imagine Just bringing parts to a line

00:04:11: Oh which totally mirrors the US market.

00:04:15: Michael Beers shared that GXO is expanding agility robotics digit deployments to new sites

00:04:20: and

00:04:21: they only did after the robots successfully moved one hundred thousand totes in Georgia.

00:04:24: Wow, a hundred thousands?

00:04:26: That's massive stress test for

00:04:28: hardware!

00:04:28: And Daniel Diaz noted that AGCO was deploying digits for parts handling by twenty-twenty seven.

00:04:34: so beach head everywhere.

00:04:36: just logistics.

00:04:37: Just moving things from A to B Makes sense, right?

00:04:40: It requires far less micro dexterity.

00:04:43: Exactly.

00:04:43: But you know.

00:04:44: if the hardware can reliably move a heavy tote but fails half-the time on nuanced kidding what's actually holding them back?

00:04:51: Well it's not that physical hardware anymore

00:04:53: Right...it

00:04:54: is the brain or data feeding the brain.

00:04:57: Mark Wallace and Naresh Redikatha have both pointed this out.

00:05:01: The ecosystem has finally realized that physical AI lacks common crawl

00:05:06: Oh!

00:05:07: And for anyone listening who isn't super deep into software, Common Crawl is like the open web data that trains chat GPT.

00:05:13: Exactly!

00:05:14: Billions of text pages just sitting there.

00:05:17: but robots can learn physical manipulation from reading texts.

00:05:20: No they need spatial reasoning Right.

00:05:23: The real mode isn't the robot's joints anymore it's the data engine capturing egocentric video

00:05:30: Which is the camera feed from the Robots own eyes right?

00:05:32: Yeah

00:05:33: Capturing and structuring in training ready tokens

00:05:36: Okay, wait.

00:05:37: How does a video feed become a token?

00:05:39: I mean in the text model of token as just part of word.

00:05:42: It's super complex.

00:05:44: Basically it's mapping visual state like seeing slipping gear to a three D motor command.

00:05:50: Oh wow.

00:05:51: so translating visual chaos into physical torque

00:05:54: Exactly and without millions of hours of that data they just freeze.

00:05:59: Aaron Prather brought really sobering reality check on this from the Robo Coliseum benchmarked Singapore.

00:06:05: oh i saw

00:06:06: Nebran, what?

00:06:06: Twenty thousand physical trials.

00:06:08: Yeah

00:06:09: And on easier tasks success hovered around thirty point seven percent.

00:06:14: Okay not great but not terrible

00:06:16: right.

00:06:16: But when they moved to the hardest most dynamic tasks The success rate plummeted to zero point.

00:06:22: four percent.

00:06:23: Zero point four.

00:06:25: I mean statistically that's basically zero.

00:06:27: yeah It's

00:06:28: like trying to teach someone a ride a bike by locking them in a classroom with a physics textbook.

00:06:32: yes perfect analogy

00:06:34: Right Like they might know the math of gravity perfectly, but the second They hit real gravel and lose friction.

00:06:40: The whole system crashes?

00:06:41: They don't know how to physically respond

00:06:43: exactly.

00:06:44: Cal no has posted something that totally backs this up.

00:06:46: He noted that these vision language action models or VLA's.

00:06:51: they score up to ninety eight percent in simulation.

00:06:53: right In a perfect digital world?

00:06:55: Yeah.

00:06:56: But if you put them in a real robot And just move the camera angle slightly Or reword the instruction success drops too Fifty-one percent.

00:07:05: Because the simulation just can't account for every single shadow or millimeter of lens distortion?

00:07:10: Exactly, which is why real world failure data Is the biggest bottleneck

00:07:14: it really does.

00:07:15: by the way if you're listening and you want to stay on top Of how these AI models are evolving week-to-week Definitely make sure to subscribe so you catch our future additions.

00:07:23: Yeah You don't wanna miss those.

00:07:24: So while Western companies are meticulously building benchmarks to solve this data gap China is taking a completely different approach.

00:07:31: Oh

00:07:31: right They aren't waiting around for perfect lab data, they are just brute-forcing the learning loop through massive volume.

00:07:38: Gene Nang did a great analysis on this.

00:07:41: China has transformed from being just fast follower to most dynamic market of embodied intelligence Driven by

00:07:49: their unmatched supply chain and government backing right?

00:07:52: Yeah

00:07:52: an extremely rapid iteration.

00:07:54: Nikhil Kumar shares some wild shipment figures.

00:07:58: Well, US based figure has a couple of robots at BMW.

00:08:03: Right?

00:08:03: But China's Unitry and Agibot?

00:08:06: they have shipped over five thousand units each.

00:08:08: Wait!

00:08:09: Five thousand each.

00:08:10: That is incredible scale.

00:08:12: And you think about how that solves the token problem.

00:08:15: right.

00:08:15: Exactly Because if put five thousand robots out there They are going to fail constantly.

00:08:20: Oh absolutely Uneven floors weird lighting

00:08:23: but every single failure captures that edge case data back to the foundation model.

00:08:29: So they are violently accelerating The Learning Loop compared to like a US lab,

00:08:34: right?

00:08:35: But you know there's a crucial counterweight To this whole narrative.

00:08:38: Henry J actually visited CMAT Asia and tour Chinese logistics facilities.

00:08:42: And what did he see?

00:08:44: despite those massive humanoid shipment numbers He noted that the humanoids Are still really slow in extensive training.

00:08:51: Really so They aren't just taking over immediately.

00:08:54: The real scale in China right now is an autonomous delivery and traditional warehouse automation.

00:08:59: Oh, okay

00:09:00: You pointed out they are processing mind-boggling volumes like fifteen million parcels a day.

00:09:06: Fifteen million?

00:09:07: Yeah

00:09:08: Which brings us to a vital point for the manufacturing professionals listening today.

00:09:12: Right you don't actually need a bipedal humanoid To solve your immediate labor shortage.

00:09:16: No ya don't!

00:09:17: The highest immediate return on investment comes from standardizing cheaper hardware.

00:09:22: Well, boring tasks.

00:09:24: Yes.

00:09:25: Boring ROI.

00:09:26: I love that.

00:09:28: Danny Bernstein made a brilliant point about this.

00:09:30: What was it?

00:09:31: He said outdoor and rugged robotics.

00:09:33: modes are completely shifting because the hardware is just getting incredibly cheap.

00:09:37: Oh right

00:09:38: Like you can buy an NVIDIA Jetson module for two hundred forty nine dollars.

00:09:42: And for anyone who doesn't know That's basically a highly capable mini computer that sits Right on the robot so It processes AI at the edge

00:09:49: Exactly!

00:09:52: not to bump into a wall.

00:09:54: The expensive parts now are integration, ruggedization and service

00:09:58: Which perfectly sets up this example from Rybin Matthew.

00:10:00: It's the perfect illustration of high ROI on totally boring tasks.

00:10:04: Oh!

00:10:05: The drone one?

00:10:05: Yeah Corvus Robotics deployed drones To scan dermatologist warehouse inventory.

00:10:11: Right

00:10:11: Not futuristic.

00:10:12: just flying up or down aisles

00:10:13: Exactly looking at barcodes.

00:10:15: But because they don't sleep They scanned entire inventory fifty-two times a year.

00:10:20: Wow And doing that freed up at one hundred twenty manual labor hours every single month.

00:10:25: That's a massive immediate return.

00:10:27: It really is it so boring, but so profitable

00:10:30: and Ralph Goldie reinforced this perfectly when talking about Sariax Cortex AI in the auto group.

00:10:35: Right

00:10:36: he said The true test of physical AI isn't some flashy demo.

00:10:40: Its weather delivers on the night shift In week twelve When nobody is filming

00:10:44: Week twelve On the nightshift.

00:10:45: that Is the True Reality Test?

00:10:48: To wrap up this theme, we have to talk about safety over flashiness.

00:10:52: Oh definitely.

00:10:53: Christoph Hommel and Hendrick Benz both noted the recent weight reduction on NeuroRobotics.

00:10:57: four net one gen-four humanoid

00:10:59: Right because in industrial settings mass is a huge safety parameter.

00:11:03: Exactly If the AI makes one of those zero point four percent errors and miscalculates friction

00:11:08: And falls over.

00:11:09: Yeah A lightweight robot Is significantly less dangerous than heavy one

00:11:13: Right.

00:11:13: Henrik Benz made the point that dancing on a stage is an easy engineering challenge!

00:11:17: Sure, The Floor Is Perfect

00:11:19: But being certifiably safe On An Unpredictable Construction Site Where Topography Changes Daily

00:11:25: That's A Nightmare

00:11:27: It Is.

00:11:27: You Need A Robot Who Won't Crush A Worker When It Inevitably Hits An Edge Case?

00:11:31: It Hasn'T Learned Yet.

00:11:33: Totally

00:11:34: Pulling All This Together The Cheap Hardware The Boring ROI It Leaves Us With A Really Provocative Thought.

00:11:41: What'S Up?

00:11:43: with foundation models becoming open source and capable brains costing less than a pressure washer.

00:11:48: Yeah, the future robotics might not actually be monopolized by massive tech giants.

00:11:54: wait really who else would monopolize it?

00:11:56: The ultimate winners could just be local farm mechanics or warehouse managers.

00:12:00: oh wow

00:12:02: yeah people who just buy cheap off-the-shelf parts to build custom robots that solve their exact operational bottlenecks.

00:12:10: That is wildly fascinating thought.

00:12:12: to leave on, just the total democratization of automation.

00:12:17: Well if you enjoyed this episode new episodes drop every two weeks.

00:12:20: also check out our other editions on smart manufacturing and industrial AI digital construction and connected tools and equipment.

00:12:28: yeah

00:12:28: definitely check those out.

00:12:29: thank you so much for joining us on this deep dive and don't forget to hit that subscribe button.

00:12:34: see next time.