Data, Memory, Gyms and Models - NEURA's strategy

00:00:04: Robotic in the industry, the podcast with Helmut Schmidt and Robert Weber.

00:00:10: So have everybody and welcome to a new episode of our Industrial AI Podcast!

00:00:15: My name is Robert Weber and it's a pleasure to talk too!

00:00:18: Peter at Sieberg.

00:00:19: good morning Robert.

00:00:20: long time no see no hear.

00:00:23: Good Morning, good afternoon Good evening to all of you, dear listeners.

00:00:27: Great To Be Back Peter!

00:00:29: So we have a very interesting episode in the main part.

00:00:32: but before We start with The Main Part let's do A Quick Newspot.

00:00:35: What Do You Have?

00:00:37: I'm gonna talk about April First Fools' Day.

00:00:40: Okay

00:00:41: My Latest Hypothesis i think it is going to be the end Of April first Fool's day.

00:00:47: In a World Full of Hype and Fake How Could We Ever Survive with or without April first.

00:00:57: Not sure that you're ready a week or two back now, but did your experience one of those incidents?

00:01:02: April First Fool's Day Robert?

00:01:04: Normally my kids are involved in this.

00:01:08: what at this time?

00:01:09: they missed it.

00:01:10: so its quite

00:01:11: interesting.

00:01:12: That is almost already going to support my hypothesis.

00:01:16: But then there our kids.

00:01:17: I'm gonna share one of them.

00:01:19: So This Is From SAP Digital Manufacturing.

00:01:23: They say introducing SAP Digital Manufacturing, agentic AI edition.

00:01:29: Now our AI agents will run your factory for you full autonomous production scheduling self-repairing machines robotic work workforce management.

00:01:40: And it's one of those, you know AI generated images and use came over.

00:01:44: Do you think?

00:01:45: Yeah sure why not?

00:01:48: everybody can say anything right.

00:01:50: but I didn't see anything weird about until maybe at the bottom assess coming April first But still i hadn't thought of anything weird.

00:02:00: You know, what is wrong about robotic workforce management?

00:02:03: What's wrong with self repairing machines?

00:02:07: you know.

00:02:07: so until somebody was suggesting and I think they there Sorry, I didn't see the name if you are hearing this yourself.

00:02:16: You were saying oh okay i'm gonna laugh about This myself.

00:02:20: we SAP We were making fun of it and I thought well Are you really?

00:02:24: Making fun Of what other people are doing already that's What.

00:02:28: That was a bit strange I think.

00:02:30: so bottom line still is you know are we Gonna?

00:02:33: Is any day In the near future, going to be an April first full day because you know it's so difficult.

00:02:39: To skin through and recognize what is truce from reality?

00:02:45: Absolutely that's interesting.

00:02:46: yeah Because I read in article i think in The German Huntersblatt That SAP Is now working on a new AI strategy.

00:02:55: Let see What does that mean.

00:02:56: yet Yeah, you go.

00:02:58: But it looks like at least one of them is not yet so fully convinced about their strategy.

00:03:04: they're making a April Fool's Day out of this.

00:03:08: I didn't go into the detail i only saw.

00:03:11: maybe You can confirm that?

00:03:13: I don't want to put wrong rumors but there was a suggestion.

00:03:18: There was an headline somewhere that their stock had gone down rather strongly, but again I want to be very careful with this because i did not check.

00:03:28: Every now and what I've been doing is checking the NVIDIA stock for a year every now-and then just because I know where it came from.

00:03:37: that's even not specifically And somebody is saying the market's going down or up, whatever.

00:03:55: I just look at the Nvidia and then I know okay if it still around The number that I have in my brain Then everything is fine.

00:04:02: If its really below Or above?

00:04:04: At last time It was strongly above again.

00:04:07: That's a number i've been following almost like for year.

00:04:09: So what you reading article?

00:04:12: Was that correct?

00:04:15: Yeah, I think the most valuable company in the German stock market is now Siemens.

00:04:21: I

00:04:21: saw that.

00:04:22: yeah

00:04:23: SAP was the lead last year but they lost traction i think and for sure when you see what's happening with Anthropiq.

00:04:33: we discussed this last time.

00:04:34: They want to go into the tabular thing right?

00:04:37: Tabular data That's a SAP domain so they need to act.

00:04:42: I think Oh Seph is next.

00:04:45: Zepp is reaching to the podcast.

00:04:47: Okay,

00:04:47: Zepp?

00:04:48: Hello!

00:04:50: It's me, Zeppe.

00:04:53: Cool podcast you're doing.

00:04:55: Thanks Zepp so let's go on Peter.

00:04:57: yeah

00:04:58: I was just writing down Zepp's name.

00:05:00: i was gonna check with you anyway more off of the podcast but... ...I was looking at people coming our AI in Alps and I wrote down that Zepp as number one.

00:05:14: I always do that.

00:05:15: As you know, when there's new names at eye people that maybe don't know what put a picture next to them and zap is number one without a picture because i knows from the very beginning.

00:05:27: so great!

00:05:27: Great to see and hear that jep is sorry that zapp with you close to coming us through the alps

00:05:37: yeah...I would like to briefly look back our episode on leaderboards last week.

00:05:44: I was surprised.

00:05:45: i received a lot of feedback and uh, A great deal of support for the researchers from pata born And i received an email and this is very interesting.

00:05:55: It's written hello robert?

00:05:56: I wasn't aware Of this topic before.

00:05:59: it's great that you also shed light on what might seem like niche topics Because leaderboards are really niche topics but i think it's a good way to describe our podcast.

00:06:10: right we also shed light on niche topics, not only a genetic but also niche topic when it comes to research.

00:06:16: When it comes industry eye and I'm really happy about this feedback because It's all about what we are doing every week.

00:06:24: Thanks a lot.

00:06:24: Yeah

00:06:25: Very good.

00:06:26: So that is specifically the is at the time series leaderboarder?

00:06:31: yeah The TS arena.

00:06:33: exactly

00:06:34: Because and i do still look.

00:06:37: And what is the name of that other leaderboard?

00:06:39: You know, I've been following Peter.

00:06:40: It's

00:06:40: Feth Bench and Gift Eval.

00:06:43: Yeah

00:06:43: right

00:06:44: and LLM Arena That's

00:06:45: the one we're in beginning here and i think there was a peter.

00:06:48: What's peter?

00:06:49: peter sautek?

00:06:50: well no what Is his Name?

00:06:51: yeah yeah yeah

00:06:52: and peter has Been There and i Think he joined him now.

00:06:55: so thats where i don't really Actively Involved.

00:06:59: i thought it Was A great Idea from Beginning Where They did and Including Now The Time Series Leaderboard as Well.

00:07:04: they are Of Course They Are Very

00:07:07: Yeah, but it's a niche topic.

00:07:08: It's not like let's talk about the new agentic workflow by I don't know who But it's good that we also shed light on niche topics for our industry guys.

00:07:18: Yep

00:07:19: most certainly and II think i did suggest in the past That I'm being thrown back like whatever is up thirty years now maybe Not as much, but Maybe it Is.

00:07:29: you Know there was The time of Looking what processors Dude There Was A Time I started with.

00:07:34: You know forty six was just in the market.

00:07:38: people who have heard about Forty-six that was just before Pentium.

00:07:41: right when I see two three numbers four eight and six my brain Brings me back immediately to their time, right?

00:07:48: And that was a time where everybody.

00:07:51: I think it wasn't like as if Davis magazine and Everybody was talking about leaderboards on how compare processes and that time, it was a time of Motorola.

00:08:02: Motorola was the very big number one in graphics you know there wasn't.

00:08:07: no.

00:08:08: am I correct?

00:08:09: In saying there was no NVIDIA ooh i want to be careful but at least not under the name of NVIDI.

00:08:15: that was...that was not.

00:08:16: they were maybe somewhere on Avery specific graphics leaderboard But I mean more in general.

00:08:21: so how do?

00:08:25: And the new processes that were coming all dependent.

00:08:28: That was very, very similar to what today is happening with our LLMs I guess right?

00:08:36: But

00:08:36: this was important for your industry.

00:08:38: so you are part of this industry and looking at the leaderboards but for consumers or customers.

00:08:46: it wasn't as important

00:08:48: in the end not well i mean we At Intel at that time we did this intel inside.

00:08:56: That was a dad-time for the consumer.

00:08:59: We were telling the consumer you don't it doesn't matter whatever PC You're gonna buy as long as there's an Intel inside.

00:09:05: now, that's and of course a big Branding thing.

00:09:09: And I think at this time we are exactly the same phase where the decision makers of industrial AI who cannot wait another year or they don't want to, I mean.

00:09:20: They already started and wanna continue doing things.

00:09:22: They need to be able have a good feeling for

00:09:26: what

00:09:26: are the different solutions giving me?

00:09:29: And then at some point in time we're gonna make their decision.

00:09:32: one is going say... that has been the best and maybe, if next week there's going to be another one.

00:09:38: The best I'm gonna take is somebody else who has different reasons for choosing a number two or three but still very good.

00:09:47: And in that perspective, it is the same as the IT decision makers maybe twenty-five years ago and still today.

00:09:56: But but today more... In a market has very matured.

00:10:02: there's other reasons why you're fixed already with certain architecture the market at which we're in with AI, with LLMs.

00:10:12: With all of things taking place is very useful that you can go and look to see who's doing more?

00:10:19: Exactly!

00:10:20: So what else do you have Peter?

00:10:22: For me, the next topic is about getting deeper into the meaning of industrial and great AI.

00:10:31: There's Steven Yates.

00:10:32: he started a discussion.

00:10:34: He says PLC has got one thing right that the edge AI industry Is unlearning which is about keeping authority local.

00:10:43: now you know HAI has been one of my, you know kind things for a very long time.

00:10:49: I think we started with H AI?

00:10:54: Actually yes me while still at Softing and where were doing the industrial AI We start it with H at that time.

00:11:01: now somewhere in discussion there's guy called Kiran Zebach.

00:11:05: he summarizes this question nicely says AI can be probabilistic But the local trust and the enforcement boundary cannot.

00:11:15: that was beautiful said when I read it And before

00:11:19: in

00:11:19: my words, and I want to share those.

00:11:22: I said you know assuming AI adds value right because if not there's no need We're gonna deploy a ride.

00:11:28: its execution still can't be less than industrial grade.

00:11:33: user term does so great at.

00:11:35: We always say Boris from Siemens has come up with, right?

00:11:41: E.g.,

00:11:41: by means of a deterministic government architecture.

00:11:46: I want to step back little bit.

00:11:48: It's very interesting i believe To understand where Steven is coming from.

00:11:52: Stephen He started his career as hardware engineer and it almost like getting Back to what we were just talking about Designing PLC CPUs at GE Fanook.

00:12:04: you know, there was another discussion that just comes to my brain right now.

00:12:09: It's always about IT OT.

00:12:11: this is a little bit separate.

00:12:12: I come back but it was exactly the same topic.

00:12:15: and then i was looking back because i said IT & OT have been seperate since they were developed.

00:12:24: And Dan That Was Very Very Interesting.

00:12:26: You Know The First PLC Was Actually Before We Had A PC.

00:12:33: It was the no, I'm not sure which one it is.

00:12:36: I think what's the Modbus?

00:12:38: Oh and i'm not so much on us but it was in nineteen sixty eight we had already And they were Not like on a microprocessor But They Were On Different Subsets.

00:12:49: But Its Very Interesting To Know That At The Time That Like Intel Was Doing The I Just Said Forty Six.

00:12:56: Then There Was A Four Thousand Four Processor.

00:12:59: other people in a different part of this world were designing their PLC CPUs.

00:13:06: In this case, so Stephen was one of those guys and he says the PLC set and it's running its ladder logic locally, whether the network is up or not.

00:13:22: And that is the big point right?

00:13:24: Even if later on there was a network and it was down mean to the PLC knew exactly what was happening rights so they authority never left the cabinet.

00:13:33: now he says Now If we're gonna go with H AI then we're going to assume that the connectivity between what is happening on the side of, has been the PLC and all my elements in a network.

00:13:52: It's gonna sit somewhere else right?

00:13:53: And he says his point this, The Edge needs become its own governments domain which In the past, the PRC as always bit now.

00:14:02: maybe it was Sounds a little bit complex maybe, but I think it's very important topic.

00:14:08: The idea being that if today something happens in your production environment, right?

00:14:16: Then the PLC knows about it.

00:14:18: Knows that something is wrong and makes a decision.

00:14:22: if with edge AI Something happens at the Edge And your network Is not connected to a central government then maybe The Edge are still going to continue To work.

00:14:32: That's I think that's the basic idea in the Central Government Organization is Not aware of You know, he's not aware that something is happening.

00:14:41: So I think it's a very important topic Sir R. Steven who was the CEO of Fedorand to come and share his vision in a podcast session about what he calls connected autonomy for H. A.

00:14:56: Very interesting concept

00:14:58: are very important as well.

00:14:59: where did you find it?

00:15:00: on LinkedIn?

00:15:01: Yeah, yeah.

00:15:02: Typically typically those things do happen on LinkedIn.

00:15:06: Very interesting for me.

00:15:08: that's the main place where I get to information you're not sure?

00:15:12: You'll go more to conferences.

00:15:14: Conferences!

00:15:15: You went through robotics conference i believe exactly and over.

00:15:20: we've been in the past so they...I'm no good reader this year but did see a lot of A lot of what.

00:15:31: should I say, messages with regards to you know coming to an over.

00:15:35: please join me.

00:15:35: But so was just saying that LinkedIn is the major source for What I see happening in the industrial space?

00:15:44: Yeah

00:15:45: It's very interesting The concept.

00:15:46: i'm really looking forward To the episode With him because he Should explain what does it Really mean?

00:15:51: yeah

00:15:52: Yeah, exactly.

00:15:53: And as I said it's a guy coming from designing but at the same time you can expect questions for me to him and say what about the ladder logic was then invented?

00:16:06: At the same times together with the first PLC right... That is where everybody now are saying that i am very completely open to this.

00:16:17: new young people interested, studying learning but at the same time you know all a generation like himself who now sees the capabilities of AI and saying are we gonna stick to do?

00:16:30: We need to stick to the ladder logic.

00:16:32: And of course that's a huge discussion with you know ninety nine percent of the people or hundred percent other people having learned And then now people saying, well I don't need to let a logic anymore.

00:16:47: But at least i'm saying we need to be in the end where the robber hits the road... ...I believe that is kind of the discussion on what we call industrial gaits.

00:16:57: You know?

00:16:57: We have to be deterministic!

00:16:59: We cannot be probabilistic when you say okay one time it works and other times not.

00:17:05: So I think potential for AI is do all kinds things very specifically, maybe around the generative.

00:17:14: So the Jakub Tomczak side of AI will say it can help us to marvel as great things but I believe that has to take place in the end when we do the execution with a deterministic way within certain boundaries.

00:17:30: What else you have Peter?

00:17:32: Two more.

00:17:34: Yeah, one is in April.

00:17:35: One but it's not an April.

00:17:36: one fools message at least I believe.

00:17:39: not you tell me Munich based agile robots.

00:17:42: stay close to acquisition of the T-Sync group automation engineering.

00:17:46: now that had already been announced.

00:17:48: end of twenty twenty five.

00:17:49: Not sure if we announce it It wasn't in my brain specifically.

00:17:54: Now what is interest?

00:17:54: was

00:17:55: robotics?

00:17:55: Yeah, so it's a decent group automation engineering.

00:17:58: more than seventy five years Engineering including robot systems.

00:18:03: six hundred fifty engineers tank global sites and agile robots as the seven year old DLR stands for Deutsche Luft und Raumfahrt German Aerospace Center spin out two now thousand employees.

00:18:17: Is that correct?

00:18:17: Yeah, the DRR.

00:18:19: I couldn't believe they are as big focus on advanced robotics to make AI manufacturing smarter more flexible and day having a quality former Franco Robotics.

00:18:30: right now following his acquisition The renamed now called Krauser automation together with HR robots They will shape next generation of physical AI in industrial production.

00:18:43: You know, that's exactly what everybody is doing.

00:18:47: I don't mean it in a negative sense but the physical AI since Janssen has brought it up when did he?

00:18:53: half year ago?

00:18:54: Was it last year or no?

00:18:56: Oh yeah Yeah Last Year i think.

00:18:58: But its amazing!

00:18:59: Really means that If you look at where you and me started seven eight years ago And now Everybody Doing Industrial AI And then The Physical AI Industrial Production Now, the second thing related to Agile robots.

00:19:17: I saw what was their strategic partnership with Google DeepMind and that's kind of what we talked about two weeks ago because in our news section We put up the hypothesis

00:19:31: about Google intrinsic.

00:19:32: Yeah,

00:19:33: we put up to hypothesis and I've seen that again somewhere else a discussion around it.

00:19:38: you know is Google going to become yet another robotics company?

00:19:43: And I said i don't believe It!

00:19:46: I Believe they're Going To Become The Android for Robotics Platform Layer.

00:19:53: at That Time that we refer to this intrinsic, they had some kind of activity up on their website.

00:20:01: And this seems to be like another announcement supporting.

00:20:06: and then there were other people saying oh yeah anybody... There's so many companies who are trying put the platform layer here?

00:20:15: Then I was asking somebody else suggesting well Who Are Their Potential Contenders?

00:20:20: AWS?

00:20:22: Did somebody suggest Tesla as well?

00:20:24: And then I and i thought like, Well okay.

00:20:27: If there's one company who is capable of really doing it Who has done at once with Android right for the handheld?

00:20:35: What is it?

00:20:35: Ninety percent Of the worldwide market?

00:20:38: My feeling was that they could do It!

00:20:40: And i Was just going to refer To this As a potential you know Confirmation in The direction Not sure how many partners they have in the meantime that they could be doing this Android for robots layer.

00:20:55: I'm not sure what you think of that idea?

00:20:57: Yeah, yeah, i will talk to Lawrence.

00:21:00: at the end of April he was part of our AI and Amsterdam group.

00:21:04: He's a R&D guy from Agile And we'll do an episode on it.

00:21:08: It is perfect topic because in the main part also focusing on robotics.

00:21:15: But we come later to the main part, but very interesting.

00:21:18: I will ask Lawrence if he can talk a little bit about deep bind cooperation in future?

00:21:24: Okay yeah looking forward who knows it doesn't have mean that just because they did their hand tell market in the robotics market quite some time.

00:21:37: And so yeah, it's just my feeling that... That is what they are going to be trying to do.

00:21:44: The final one I have here maybe before you then share with us this thing we're talking about.

00:21:48: on our main section There's James Booneak who launched SNAP PLC.

00:21:53: He says take a picture of your control cabinet and AI automatically generates the entire PLC program.

00:22:01: Very impressive!

00:22:01: Advanced features reconstructs ladder logic.

00:22:05: There it is again, from the way the wires feel emotionally connected.

00:22:11: listen what what the verbiage is right?

00:22:14: It detects undocumented logic written by contractors in two thousand seven.

00:22:20: wow this feels like magic!

00:22:22: Right?!

00:22:22: It predicts which I owe point the electrician meant to land.

00:22:28: It gets better all the time.

00:22:30: And it also, it recreates the original programmers thought process including panic.

00:22:36: I'm not going to say more is just no will not because this was actually an April one false joke in case you hadn't learned yet.

00:22:49: i think if you haven't been around engineers that are being called on site, and all they find is exactly this ladder logic.

00:23:08: And somebody touched the running system ten years ago... ...and now for whatever reason something has gone wrong!

00:23:15: You do not know what was wrong!

00:23:19: The programmer needs to go into it.

00:23:21: there's no documentation or nothing.

00:23:24: But again, that's the thing I will talk to Stephen about as well.

00:23:29: It is finding out... Do we need to go into this ladder?

00:23:34: Which elements?

00:23:36: maybe just round it off around April.

00:23:41: first.

00:23:42: which of these elements are more now to turn at a hundred eighty degrees around almost like these things too.

00:23:50: good be true.

00:23:52: But some points of these, I do believe will become true in one or two years.

00:24:00: Really!

00:24:01: And this is meant to just make fun and joke... ...and the frustration of the electrician.

00:24:07: but....I do believe that AI will help us too you know?

00:24:12: And i haven't been smoking.

00:24:14: I do believe that in one or two years, certain elements of this.

00:24:18: Yes exactly!

00:24:19: That's what AI is going to be helping us and i think the central question for this nevertheless it like are we gonna then stick to the things that we have been doing?

00:24:29: Or Are We Going To Be Doing Things A Different Way?

00:24:34: Maybe even so with AI on top Of It?

00:24:37: Let The AI Take Care.

00:24:40: I understand that's a very, very... What should i say?

00:24:45: Dangerous maybe.

00:24:48: Saying because there is going to be now many of you saying well Peter what are your talking about?

00:24:54: You can't do it but That has been said about coding years ago and you recall And everybody kind of doing one way or the other.

00:25:04: So let see if certain elements April's first Fool's Day joke is gonna and nevertheless going to become capable in the near future.

00:25:15: Robert, what

00:25:16: are you?

00:25:16: It's a better dope than The SAP Joke!

00:25:18: Yeah

00:25:18: exactly.

00:25:20: So let's move through the main part because I did an interview with Jonas from Neura And here's one thing i learned From the interview Because he said everything we do In robotics will require Now, a very important new forms of memory and temporal dependency in the future.

00:25:42: And this was very interesting.

00:25:44: so New Forms Of Memory and Temporal Dependency Very Interesting Episode.

00:25:48: I think we didn't talk too much about Neura.

00:25:51: We talked A little bit About Neura & Their Ideas But we Talked About Models World Models new forms of architectures.

00:25:59: Very, very interesting episode by Nora and by Jonas.

00:26:03: Thanks a lot, Jonas!

00:26:04: It was a pleasure.

00:26:04: And if he talks memory does it mean like physical hardware?

00:26:09: Or did you talk about the more logical structure of memory?

00:26:13: because that's the ladder... If its'a ladder-it brings me into without knowing details.

00:26:18: but like LSTM XLSTM I think thats different ways to deal with memory as well.

00:26:23: i believe right

00:26:24: Exactly.

00:26:24: From my personal perspective and when I talk to the robotics companies, we will see a revival that comes through recurrent neural networks.

00:26:33: so new form of kind of RNNs maybe because memory topic is very crucial in the future.

00:26:43: Okay, looking forward to.

00:26:45: I mean you didn't talk about them but maybe But i'm sure.

00:26:47: You did asking about his company.

00:26:49: probably yeah exactly because Neura Nura kind of refers almost like two To our brains and your neurological?

00:26:59: Yeah not sure if that is correct, but then they gave themselves their names.

00:27:03: so it this may be.

00:27:06: That's why he or you has been talking about the memory.

00:27:11: Maybe you're always going to be careful, not our human brain works because the people who know about that I would say was still very different from how algorithms work.

00:27:21: Okay looking forward to listening.

00:27:24: Peter

00:27:24: it was a pleasure.

00:27:25: thanks so much!

00:27:26: Robert talk soon again and dear listeners hope your gonna enjoy the main section as we already enjoyed our news.

00:27:36: have you with us.

00:27:40: My name is Rod Weber and my guest today is Jonas Messner.

00:27:44: Jonas, welcome to the podcast!

00:27:45: Thanks for the invitation.

00:27:47: nice to meet you.

00:27:48: So my name is Jonas messener.

00:27:50: I'm head of AI here at Neuro Robotics.

00:27:52: Yeah i've been in AI pretty much my whole professional life have worked in automotive For five years bringing AI into various products that are out there on the streets And then, yeah here at Nura first took over the gyms and built a gym.

00:28:08: But then in the meantime also it took over The role of head-of artificial intelligence overall.

00:28:13: so essentially my responsibility is everything from data to training models To foundation models too.

00:28:21: deploy these on our robots.

00:28:23: Okay that's interesting.

00:28:25: you are the right person to talk today.

00:28:27: we don't want just to talk about robots, but more about your ecosystem.

00:28:32: Gym architecture models foundation models.

00:28:35: what are these gyms?

00:28:37: Can you give us explain as What Are These Gyms?

00:28:41: So essentially the gyms are addressing The

00:28:46: big

00:28:46: challenge of physical AI which is scarcity Of data.

00:28:51: Essentially with the internet has been two large language models like large amounts of text and images.

00:28:58: This is not there for robotics or physical AI, because we're missing the real world interaction and how humans interact in a sense of touch.

00:29:09: Or also audio or three-d vision all that isn't present on the internet.

00:29:15: So essentially what the neurogym is about?

00:29:17: We collect large amounts of such multimodal physical AI data Such as we can even build Real-world physical AI.

00:29:27: So this is what the neuro gym is all about.

00:29:30: so you use your robots in The Gym to produce data or am I

00:29:34: wrong?

00:29:35: Yes, so essentially In our gyms we have hundreds of robots acting in real world environments.

00:29:42: So the gym.

00:29:43: oftentimes when we talk about gyms people think okay This is simulation.

00:29:47: no it's actually real World Physical Infrastructure.

00:29:51: so physical buildings tele-operate our robots to collect data in real world tasks.

00:29:57: So for example, you could think about logistics task where a robot is sorting packages or You can think about an industrial task were the robot is assembling parts and we're Operating these robots at the gym infrastructure To then get the data out and train models on it?

00:30:18: Because scaling.

00:30:20: we learned that maybe the time of scaling is over when it comes to LLM, right?

00:30:25: So are there combination where they come to simulation and using data or really only a data-driven approach.

00:30:34: Simulation also plays an essential role but what you see for interacting in physical world like fine grant manipulation the real world data is simply irreplaceable.

00:30:49: But we're having a multi-fold approach to our data strategy, so it's about simulation.

00:30:54: It's about the gym datas of the real world.

00:30:56: physical data is also about a special concept that we have in the context of equipping humans with data suits and partnerships with bigger corporations.

00:31:05: but its like a multifold approach.

00:31:07: two data collection.

00:31:09: can you share little bit about your approach when comes to collecting data in factories?

00:31:13: or was who are partnering to collect data?

00:31:17: Yeah So to collect data with partners, there are multiple partners that we're partnering within a big way.

00:31:26: One of them or two of the other public is Bosch and Schaeffler.

00:31:30: so essentially what were doing in such partnerships?

00:31:34: on one hand these companies using the gym they use it for collecting data from real robots but since then is not the solution to do everything, so there are more efficient ways of data collection.

00:31:48: We're actually also equipping workers with what we call DataSuit.

00:31:54: So essentially in factories they get a sensor on their body or sensors.

00:31:59: so cameras tactile gloves audio and all these things that our robot has such that during regular eight hour shift can collect data that we can use for training.

00:32:13: So this is a very efficient approach to get large amounts of data from real-world factory environments.

00:32:21: and the interesting thing about that, such data doesn't exist anywhere on the internet.

00:32:26: Of course those companies are careful who they give their data too.

00:32:31: We're in lucky position with strong partnerships with these companies which essentially gives us access tremendous data source.

00:32:42: Absolutely, but

00:32:43: what

00:32:44: do you do afterwards when you receive the data?

00:32:47: Do you need the annotation of the data or What is a process then?

00:32:52: so we're mostly working in this context with self supervised approaches.

00:32:58: So were trying to reduce the amount of manual labeling by yeah are two minimum amounts.

00:33:07: Having been in larger corporations, I know that labeling cost can be a significant factor and such projects.

00:33:14: And we're trying to use on the one hand self-supervised approaches but also ground truth sensory around humans as well as robots in our gyms which essentially is overlooking the entire process.

00:33:29: That ground proof sensory can then be used together with advanced AI models like large scale vision language models to auto annotate the data.

00:33:39: So, the amount of manual annotation that we do is limited to like really minor quality assurance and you know?

00:33:49: Like they're really tough cases where You still need human annotation but it's a small amounts that we manually label.

00:34:02: That's exactly the approach that we're taking.

00:34:05: So while currently most other players are really focusing on solving everything all at once, by themselves which you can imagine I mean physical AI is huge.

00:34:17: right as thousands and thousands of tasks out there have to be solved to make physical AI real.

00:34:23: We're taking a very different approach than most others here where essentially in away crowdsourcing All these data to learn the task.

00:34:33: So, the gym is really meant not for us to go in there and train everything all by ourselves but bring companies into the gyms.

00:34:43: so we're bringing huge companies from logistics from industrial service like from all kinds of areas.

00:34:50: they are going collect the data for their tasks and then also via the gym infrastructure, so that Jim is not just physical infrastructure but also huge cloud infrastructure behind to do.

00:35:02: The whole data pipelines model training and deployment To enable these companies to train.

00:35:09: they're tasked in with that actually Do the whole scaling of physical AI?

00:35:14: How do you convince them to share the data on your gym?

00:35:19: So We see that for a fact, everyone wants to

00:35:23: go

00:35:24: into the gym.

00:35:24: So when we announced that concept actually were drowning in requests to enter the gym.

00:35:32: so because of these companies being all under lots pressure as well make their factories more efficient and with our gym infrastructure essentially providing them place where they can train these robots They are all excited about this concept that they want to actually enter the gym such as it can even train robots.

00:35:55: I mean, building something like a gym just every company by themselves doesn't make sense right?

00:36:01: This is where the whole thing then again does not scale because if hundreds of companies all buy themselves and build such an infrastructure It's very costly but we built once for everyone.

00:36:15: This is where the real scaling happens and whether real efficiency happens because everyone can use very similar pipelines, a very similar training on data pipelines to train their tasks.

00:36:28: You mentioned Data Pipelines Training Pipelines.

00:36:31: yes we stop when it comes to annotation right?

00:36:33: Can you go further in process explain what happened then?

00:36:39: yeah So essentially in the gym, we have two kind of modes.

00:36:44: Yeah?

00:36:45: We're facing lots of companies that basically have very little knowledge about how such models are built and what were trying to do here is completely abstract.

00:37:02: this process data pipelines away from these can fully do the operation by themselves.

00:37:12: So they focus on the data collection piece and essentially then just feed that data into our gym infrastructure to train them all, so when magic happens in their back thats what we're focusing on.

00:37:25: but give you a little bit of details.

00:37:27: it's essential.

00:37:29: I mean your getting that data or annotating this data also in these pipelines.

00:37:34: you have curation steps can look into, okay what's actually the relevant data that we should train on?

00:37:41: Then there is training pipelines in these infrastructure.

00:37:44: There are deployment pipelines together then either back on the robot itself or also depending on use case and task to deploy towards our cloud infrastructure being the neuroverse

00:37:57: Exactly!

00:37:58: That was my question.

00:37:58: Is their connection between the gym and the neuro-verse?

00:38:03: at end

00:38:04: Absolutely, I mean the neuro gym and the neuroverse are like one in the same.

00:38:08: It's like yin and yang One doesn't exist without the other.

00:38:12: So essentially while uh The neuro gym is this data factory And the entire training pipelines to train these skills The neuroverse Is the place where you In the end build your overall applications.

00:38:26: Then you do fleet management And you oversee Essentially what Your fleet of robots is doing.

00:38:31: so they're essentially one and the same.

00:38:35: Okay, you mentioned now data?

00:38:38: You missed a little bit the whole topic when it comes to models right.

00:38:42: so which role do models play in the gym?

00:38:46: of course The Models are core piece yeah.

00:38:50: So It is always model that's trained on the Data Right.

00:38:56: And here we're approaching That Topic In A Multi-Fold Way.

00:39:02: Of course, we're using and building also on existing models that are out there from big players like NVIDIA or Google Gemini.

00:39:16: Making use of these models and tuning them to our needs is one approach.

00:39:21: but at the same time see such models today mainly being vision language action models some cases, like in the rather I would say easier cases but that in many more complex real-world interactions these models face significant limitations because think about it us as humans do we rely on vision only to execute them?

00:39:49: The real world.

00:39:49: I mean no right We are really multimodal Humans Like.

00:39:55: we have a great sense of touch which is usually very big portion interaction in the real world.

00:40:01: We have a great sense of hearing, plus also three division and what we are building by ourselves is our own physical AI foundation models to address these shortcomings.

00:40:13: that model today out there considered state-of-the-art.

00:40:19: so we're going even step further on.

00:40:24: I'll solve many more complex problems than these VLAs, this vision language action models are able to solve today.

00:40:32: But what is the problem with the VLAS?

00:40:34: Is it a memory topic or sensing topics and what's their main obstacles?

00:40:41: It's pretty much the sensor modalities.

00:40:44: so think about a task where you do not just pick up an object in some other place but fine-grain manipulation in the real world.

00:40:57: For example, you have to stick in a cable into a hole or something like that.

00:41:03: there The force feedback Or what?

00:41:05: You then feel on your sense of touch and the tactile senses under your robot hands is essential for task success.

00:41:13: without this data source Your essentially...you may still be successful In some cases but your success rate in terms of tasks success is much slower than if you make use of this data.

00:41:25: And there are similar examples for bringing in audio, where actually we have examples with some of our partners that show... You know what's this welding topic?

00:41:38: Where the best welders in the world don't actually weld based on vision but based on what they hear.

00:41:46: Yeah, it's actually fascinating to when we learned that because yeah.

00:41:50: That's not what you would think intuitively but its is actually the case there?

00:41:55: absolutely.

00:41:56: We had an episode on their topic I think together with a rocket science guys in Switzerland.

00:42:01: i Think they're also using sound for welding applications?

00:42:04: Yeah, I was also surprised to hear that The best welders on the earth can here.

00:42:09: What is wrong?

00:42:10: and yes thats really how us as humans interact in the real world.

00:42:16: We're a multimodal human, you could say.

00:42:19: and that's simply not covered in vision language action models.

00:42:23: And you know what?

00:42:23: The nice thing about visual language Action Models is You can train them on internet data and That's also one of the main reasons why you see all Of the big tech companies doing that because they have access to these huge amounts of data like we as well do but They don't have access multi-modal data that you need to train such systems, as I just explained them.

00:42:45: But that's where we bring in the gyms right?

00:42:48: So that's how.

00:42:52: What about memory?

00:42:53: I think when we talk about robots or the next generation of robots, We need more memory to understand what's happening in the environment.

00:43:02: In the last two minutes... Or look back at five-minute ago who i am how i handled this situation.

00:43:09: Isn't memory a topic for you?

00:43:11: Absolutely!

00:43:12: Memory is core piece in our own models as well and that actually also one of shortcomings.

00:43:19: in addition to multi modality today's vision language action models.

00:43:24: They're typically frame based, now they have no sense of temporality and this is where we are also introducing essentially a sense of memory with you know the robot knows what did it do two seconds ago or ten seconds ago?

00:43:40: Or a minute ago?

00:43:41: so absolutely that's core topic as well.

00:43:44: So coming back to gyms R&D lab for you, your customers building models and then publish the model.

00:43:53: And then all of us is that way to go?

00:43:57: Yes so it's a yeah... You could call it an R & D Lab.

00:44:01: I would much more called really training world where you can train use case realize the use case not just collect data in train but also validate actually successful and then bring it into your actual application, in your factory for example or in your logistics environment.

00:44:25: And the way you do that is... You train or collect data, you train models, validate them at the gym.

00:44:32: once you're happy with results like your ninety-nine percent success rate you can publish via Neuroverse.

00:44:40: That's another part where it becomes really magic If we talk about, for example a logistics partner that they are training right now with us like package sorting as I described earlier.

00:44:54: This skill of sorting packages is useful to many more companies and can then publish the skill via the Neuroverse or even monetize it.

00:45:08: So its also platform?

00:45:11: Exactly!

00:45:12: Okay, that's interesting.

00:45:13: You mentioned NVIDIA and Google right?

00:45:14: So they have a lot of great VLAs.

00:45:16: but you mention that there are trained on internet data or public available data.

00:45:21: now we have this amount of data.

00:45:23: your collect in the amount after building.

00:45:27: Now as you mentioned own models memory with more capabilities And you have this platform in the future.

00:45:32: for now and already have it Is is second business model Or not to sell robot also to sell let's say VLAs, models new version you approach it.

00:45:44: when the comes to VLAS made by Nora.

00:45:46: Is that a way to go?

00:45:48: Yeah I wouldn't say its second but it is like directly in parallel.

00:45:54: so of course we are selling our robots full stack hardware I mean already in parallel being sold.

00:46:04: It's a second revenue stream, you're absolutely right about that and it is usable not just for Nura's own robots but also other robots.

00:46:16: We do not believe that we solve every robot embodiment for the entire world, but there can be other robotics providers to bring in their hardware.

00:46:26: Also bringing it into the Nuraverse train their skills via the NeraJim including our own foundation models and enable also these robots from third parties.

00:46:38: So for us, the platform it's not dependent on being for Nura robots only.

00:46:43: It is an open-platform right?

00:46:44: Absolutely yes!

00:46:46: The main topic then that these models need to generalize in different robotics applications.

00:46:53: how difficult is

00:46:56: this?

00:46:58: Most of parts are actually the same across embodiments because wheeled robot or a legged robot, even just the single arm robot.

00:47:10: As long as they have similar cameras and sensors for audio, they will perceive their environment in the same way that they interact with them.

00:47:22: That's of course then slightly different between different embodiments but most of the intelligence Sensing and the thinking, how you want to solve things.

00:47:34: That's the same across all

00:47:35: embodiments.".

00:47:36: And in the end it is rather like an... You may call it an embodiment fine-tuning that enables us do a right movement of the robot but perception stack is the same?

00:47:49: We

00:47:49: will see new let's call them AI integrators who are doing this fine tuning for companies at the end as business model.

00:47:59: I think the whole thing in the gym infrastructure is there will be companies acting in this, in these gym infrastructure that maybe an entire new market for a company's to act.

00:48:13: To actually enable companies to train tasks or collect data from certain tasks to then train the models and bring these models on their robots, because it doesn't need be of course a logistics company that goes into the gym.

00:48:28: They may actually subcontract data training companies or service companies who do all this in the back but still they would use our gym infrastructure.

00:48:40: And does the gym infrastructure cost me?

00:48:43: What is it?

00:48:44: Is there a price tag on this gym infrastructure for me as customer, partner or client.

00:48:50: Yes so of course the price tag behind.

00:48:54: I cannot talk about exact number now here but what we do?

00:48:58: Essentially, especially now in the beginning if you buy a robot from Nura that comes with certain amount of training capacity and data collection capacity.

00:49:09: You can do it at the gym such as enabling your robots to train or learn their tasks.

00:49:17: Jonas, what is on your AI agenda?

00:49:25: So for us right now, it's really about the topic of scaling and scaling this gym infrastructure to enable the entire world.

00:49:35: And we're actually having already like our gyms being currently built up in Germany... ...in China also in the US & Japan plant.

00:49:44: so currently our focus is on scaling on the AI side and in terms of technology, we're also following several trends.

00:49:55: Currently many people are talking about world models.

00:49:58: that may be a completely new approach but there maybe entirely new approaches to how to solve AI.

00:50:06: The interesting thing here is our whole efforts building the pipelines in the back, they will transfer one-to-one.

00:50:20: if then suddenly a new methodology rules physical AI.

00:50:26: The data source that we're growing on the back is going to be transferred from time to time and this is also why we are focusing so much on scaling and not just worrying about what's right model for success.

00:50:37: because I mean this whole technology of foundation models was so young.

00:50:42: it has been years old.

00:50:44: It will certainly change.

00:50:45: The types of models that we see today, Will not be the type of models That would solve things in three or five years.

00:50:53: But what is going to be the same Is data.

00:50:55: This is why I focus on that data part so much And getting good high quality data and also the right amounts Of such data.

00:51:06: Even if technology advances significantly We're always having that huge source of data in the back.

00:51:14: Okay Johannes, I keep my fingers crossed for you!

00:51:17: For your company, for your gyms... ...for your new world models and VLAs.

00:51:23: Thanks a lot.

00:51:24: it was a pleasure.

00:51:25: Thank You so much.