Latest / RobTalk / What Physical AI Is Not: And Where It Works Today

Transcript

What Physical AI Is Not: And Where It Works Today · RobTalk

• 0:00
  Speaker 1: So now you have movements and we need to make decisions in between.
• 0:04
  Speaker 1: Every decision point is basically a new probability distribution, then those multiply...
• 0:09
  Speaker 0: That makes lots of sense!
• 0:10
  Speaker 0: It also makes sense that it can't give too complicated commands or too much data at the same time because just the processing doesn't work.
• 0:17
  Speaker 1: Sure….
• 0:17
  Speaker 1: And there's also energy usage right?
• 0:19
  Speaker 1: Such a cluster I talked about has maybe one hundred kilowatts.
• 0:23
  Speaker 1: We've got a bucket of ice here and Timon would you please put your hand in there for two minutes.
• 0:34
  Speaker 0: I can hardly grab them, this one is small
• 0:39
  Speaker 1: so you can differentiate.
• 0:40
  Speaker 1: now i can say okay what tasks are actually working pretty well?
• 0:44
  Speaker 1: So here for example have a pile of paper and so now if at trying to lift that piece of paper it might work But it also is actually a bit finicky and if I take it out, you know things are not as easy might fall
• 0:58
  Speaker 0: off.
• 1:02
  Speaker 0: We're back again with the new episode of Rob Talk And today we talking about something really special.
• 1:07
  Speaker 0: What's possible to day with robots?
• 1:09
  Speaker 0: Not what happening in ten years from now or everybody showing on internet?
• 1:13
  Speaker 1: No!
• 1:16
  Speaker 0: What's possible today with a Robco robot?
• 1:18
  Speaker 0: what we can do right now?
• 1:19
  Speaker 0: physical AI in your production line?
• 1:22
  Speaker 0: And if you're in CAO of your production manager, If you have AI budget Right Now and You're looking how to spend that In the physical AI space or if you are Looking To maybe get into The Physical AI Space With Your Production Or Your Factory.
• 1:35
  Speaker 0: This is the right episode for you because We'll let you know what you should think about How to set your expectations right.
• 1:43
  Speaker 0: Well, to make a project successful and not make it fail.
• 1:46
  Speaker 0: Because I think that's the biggest danger right now is not about testing it but testing it wrong so that The outcome was wrong And then you never look at again?
• 1:56
  Speaker 0: Then basically... ...the future has gone and everybody else overtakes you.
• 2:00
  Speaker 0: Clarence thank You for being here!
• 2:01
  Speaker 0: Now know you've brought some crazy appliances as well.
• 2:05
  Speaker 0: Some of them are for me too test.
• 2:06
  Speaker 0: So let us talk today.
• 2:09
  Speaker 0: Let´s get into practical
• 2:12
  Speaker 1: The Autonomous Robotics Podcast.
• 2:15
  Speaker 1: Physical AI, no theory just reality.
• 2:23
  Speaker 1: Yeah absolutely I mean the background is that obviously we've been looking at this space for a couple of years and when i came in you first build visions where you want to be three five ten years but then you execute.
• 2:41
  Speaker 1: We have gone a long way since And this is basically what I'm doing all day long, right?
• 2:47
  Speaker 1: How to make that vision real.
• 2:48
  Speaker 1: How do you build the teams how to make sure we built the right things and so on?
• 2:52
  Speaker 1: but then The other thing is that i'm doing every week Is to sit down with our sales engineers With our uh... with Our sales people and with our product managers To go through really customer case by customer case and understand how can we do this, when you probably see a picture or video of the process.
• 3:16
  Speaker 1: And then you have to think okay... How would our robots fit in here?
• 3:22
  Speaker 1: From that over months and months doing this find out certain patterns.
• 3:28
  Speaker 1: so often enough when a sales engineer now comes and says ''How should I do
• 3:32
  Speaker 1: this?''
• 3:33
  Speaker 1: ...I ask questions back And I think this is exactly what we want to cover today.
• 3:38
  Speaker 1: Some of the things that are really simple gates, how can they make your first AI-based robot project success?
• 3:48
  Speaker 0: This will be a great episode because it gives you practical advice for what... Yeah!
• 3:52
  Speaker 0: You can achieve today.
• 3:53
  Speaker 0: so if you're startup or midsize company and start with robotics where everything's at i think Yeah, tell me what's going on right now.
• 4:06
  Speaker 0: And you know I mean a lot of people are talking about this.
• 4:09
  Speaker 0: but can the robot already solve a real issue like like a plumber or something?
• 4:14
  Speaker 0: Like that.
• 4:14
  Speaker 0: so if i'm having an issue with my heating and from having issues in my pipes In My House Can I call Robot uh...and he can come fix that stuff already.
• 4:24
  Speaker 0: Or is That too far away?
• 4:26
  Speaker 1: yeah it's probably not The first problem I wouldn't want to solve Right because If You think About It Really Deeply What A Plumber Has To Do?
• 4:33
  Speaker 1: There's a ton of problem solving along the way in three D space.
• 4:37
  Speaker 1: And there is more specifically, there's a lot of crouching behind toilets and environments where things can break.
• 4:47
  Speaker 1: There's A lot of intricate tasks.
• 4:50
  Speaker 1: were you need to screw something call these dexterous tasks right because we need all the fingers um an and You Need To Yeah Really Work In Very Tight Spaces.
• 5:02
  Speaker 1: And even if you think about it, how many muscles a human body has.
• 5:08
  Speaker 1: For example lifting the shoulder is something that If You Build A Robot It's One Of The Things That You Probably Leave Out Because Its Like Not Needed For Many Things.
• 5:18
  Speaker 1: So Your Still A Lot Stiffer Than What A Human Could Do.
• 5:24
  Speaker 1: I Think Plumber As A Job Is Still Safe for A Long Time and even just being able to solve all the problems, understand what problems are there.
• 5:37
  Speaker 1: And then deal of course also with fluid stuff like that.
• 5:42
  Speaker 1: so yeah very challenging.
• 5:43
  Speaker 0: What about carrying stuff or I don't know lifting stuff?
• 5:47
  Speaker 0: That's perfect especially heavy stuff.
• 5:50
  Speaker 0: right it is perfect for robots because they're incredibly strong.
• 5:53
  Speaker 1: Yeah but i think Um, imagine a little bit.
• 5:57
  Speaker 1: It's it's like with cameras when you have the uh Like ISO hundred versus iso thousand.
• 6:04
  Speaker 1: yeah So so it's good if you stay within let's say one order of magnitude and then once You move to a different order of Magnitude that becomes difficult again And then you need maybe A different way how The system Is set up?
• 6:17
  Speaker 1: So for example If you if you Need To you know thread a needle um you Have very Very small forces and very intricate detail versus lifting a bucket is, it's very different.
• 6:31
  Speaker 1: So you probably would want to have just the different robots for that.
• 6:35
  Speaker 0: Okay?
• 6:36
  Speaker 0: Will there be one robot who does all or will we have specific robots for each task?
• 6:43
  Speaker 1: Well I think its beauty of sticking relatively close with human form factor.
• 6:51
  Speaker 1: We can a very wide range of things, but just you have to be careful in the case by cases.
• 7:02
  Speaker 0: But You've seen the news right?
• 7:03
  Speaker 0: AI is being presented as going basically off-the-rails or crazy like I can do anything now that knowledge worker could and where's the robots at?
• 7:17
  Speaker 0: so i've seen people talking about this online.
• 7:21
  Speaker 0: Robots can do everything and they will place all the human jobs as well.
• 7:24
  Speaker 0: And my question is, how far are we there?
• 7:27
  Speaker 0: Because looking at what AI on the coding front can do with our computers right now it seems like a lot of jobs are changing a lot.
• 7:39
  Speaker 0: And AI's doing a lot work in my computer already.
• 7:42
  Speaker 0: that I used to do maybe two years ago myself online.
• 7:48
  Speaker 0: It seemed like robots could be able.
• 7:51
  Speaker 0: Are we there yet?
• 7:52
  Speaker 1: Right.
• 7:53
  Speaker 1: So I think maybe what i would like to do is talk a little bit about our day-to-day work, right?
• 8:00
  Speaker 1: I started this whole thing two years ago and obviously when you start doing the stuff then your first build visions And so it could be talked about where will.
• 8:14
  Speaker 1: But the other activity that I'm doing is a meeting every week with our sales engineers and product managers, account executives.
• 8:25
  Speaker 1: And to really go through requests that come out from the field from our potential customers.
• 8:31
  Speaker 1: So after all this time during these weeks where i am getting asked what's possible there are patterns emerging Right.
• 8:41
  Speaker 1: So I can say now, meanwhile if somebody comes and says Can we do this?
• 8:45
  Speaker 1: And then I can ask back Okay have you thought about This?
• 8:50
  Speaker 1: Have You Thought About That?
• 8:52
  Speaker 1: Maybe that's the first gate that he needs to pass.
• 8:54
  Speaker 1: so i think thats something We can talk more in detail today.
• 8:58
  Speaker 0: I'd love too because for me it is like The mechanics are still gone far away right.
• 9:08
  Speaker 0: but the whole combination of sensors and making it possible, just what you also described about the shoulder.
• 9:13
  Speaker 0: You know?
• 9:14
  Speaker 0: Making this possible is like... This super hard challenge to build.
• 9:17
  Speaker 0: but then obviously the question is can we assume that a robot will move perfectly at some point?
• 9:25
  Speaker 0: or What's holding us back from seeing more robots on streets today Or seeing them in households?
• 9:32
  Speaker 0: I mean there are a lot of robots in industries.
• 9:36
  Speaker 0: We're not seeing as many robots, we thought.
• 9:38
  Speaker 0: We would see or when this is being described what's possible right now?
• 9:43
  Speaker 1: Yeah so I mean This Is What You Know.
• 9:47
  Speaker 1: Many CIOs Or Production Leads Are Asking Themselves Today As Well.
• 9:52
  Speaker 1: Maybe The Budget That People Have Today Is Not the Production Budget But An AI Budget.
• 9:59
  Speaker 1: And So Now The Question Is How Would I Spent This Budget?
• 10:05
  Speaker 1: And so I think the first thing is the mechanics.
• 10:11
  Speaker 1: By that, imagine you have a robot having an AI problem completely solved
• 10:19
  Speaker 0: and
• 10:20
  Speaker 1: would still be able to do what they want?
• 10:26
  Speaker 1: Obviously if we see some of these human-like robots looking like humans it's extremely deceptive because as humans immediately think, okay so it has all the properties of a human.
• 10:39
  Speaker 0: Exactly
• 10:39
  Speaker 1: and well in industry we wouldn't want to copy the form factor verbatim.
• 10:46
  Speaker 1: We also want be very careful putting more technology than what you want to maintain Putting more sensors then actually need And making sure that things are robust both from spare parts, but also just you know being able to give some guarantees on uptime.
• 11:12
  Speaker 1: And yeah I mean it starts with what means to sends and touch and grip whether a simple or very complicated gripping mechanism needs to be used.
• 11:30
  Speaker 0: is this the touching and gripping mechanism?
• 11:33
  Speaker 0: Is it so important?
• 11:34
  Speaker 0: because our world was made for us humans who have hands to interact with many parts of the world, that's why the robot needs those or something else.
• 11:44
  Speaker 0: The tasks are usually created for human not robots specifically.
• 11:50
  Speaker 0: probably like ten years ago when robots weren't as capable they were now production street would be catered towards automations that are provided, so there's a special hook on something and can grab it specifically.
• 12:05
  Speaker 0: I find this really fascinating because the world is basically made for humans So everything is made soft to touch.
• 12:14
  Speaker 0: For us Everythings in right height.
• 12:16
  Speaker 0: We don't hit our heads here In the ceiling.
• 12:19
  Speaker 0: If you go into car for instance A good example Car is made out of metal But then The things we interact with as human Are made softer And we dont hurt ourselves.
• 12:28
  Speaker 0: And the chairs obviously are soft and padded, everything.
• 12:31
  Speaker 0: What I find fascinating is that it's made for us to grab or lift nothing too heavy.
• 12:42
  Speaker 0: Many years ago automation lines were made.
• 12:46
  Speaker 0: if you had something... You probably have to build special hooks of special attachments in there so the robot can take them on their next station.
• 12:53
  Speaker 0: but this is changing!
• 12:57
  Speaker 0: we can build the robots more to the human world, and that makes it obviously more capable.
• 13:03
  Speaker 0: But one of the core things is that the gripper works right?
• 13:06
  Speaker 0: And uh... It's not that easy!
• 13:09
  Speaker 1: Yeah its'nt that easy because were so used for the human hands.. One other thing I'm telling in these sessions where we talk with our sales engineers To say okay imagine that AI works perfectly.
• 13:23
  Speaker 1: Can you still do the task when?
• 13:25
  Speaker 1: You don't have a human hand, but when you have just your own simple gripper and Meanwhile we We Have mock-ups of grippers that he can hold in their human hands And then you can actually take out to customer and so that now then it becomes suddenly Immediately super clear.
• 13:43
  Speaker 1: Yeah So yeah When I want to do two things here But maybe i'll first request We have a bucket of ice here.
• 13:54
  Speaker 1: And Timon, would you please put your hand in there for two minutes and then we continue?
• 14:02
  Speaker 1: While you're doing this... This looks nice!
• 14:06
  Speaker 0: Very nice.
• 14:06
  Speaker 0: It's like an ice bath from my hands.
• 14:08
  Speaker 1: Yeah it reminds me of old TV commercial.
• 14:16
  Speaker 1: So yeah one other thing is take such device with some tongs.
• 14:24
  Speaker 1: Now I can say, okay what tasks are actually working pretty well.
• 14:28
  Speaker 1: So here for example have a pile of paper.
• 14:31
  Speaker 1: and so now if trying to lift that piece of paper it might work but also is actually a bit finicky in about take-it out you know things not as easy might fall off And then with this thing It would be quite hard.
• 14:47
  Speaker 1: really make sure This works both contraptions built, you know special devices that can do it and so on.
• 14:55
  Speaker 1: So not so easy!
• 14:57
  Speaker 1: Not even with a humanoid hand I would think... And this is something that just immediately works.
• 15:05
  Speaker 1: different thing.
• 15:07
  Speaker 1: um i have here a box of pens?
• 15:09
  Speaker 1: What if?
• 15:10
  Speaker 1: what if I'm using those things to take out those
• 15:13
  Speaker 1: pens??
• 15:13
  Speaker 1: So suddenly that
• 15:14
  Speaker 1: works!!
• 15:15
  Speaker 1: I can take them all quite easily.
• 15:19
  Speaker 1: vision support, I don't need to feel anything.
• 15:21
  Speaker 1: Let's see when it drops back... Worst thing that can happen is that i pick two right.
• 15:26
  Speaker 1: so this works usually and then we have some contact-rich things.
• 15:31
  Speaker 1: So here now for example use a second hand open the can fix it or something where you think about.
• 15:38
  Speaker 1: okay?
• 15:41
  Speaker 1: How's the hand, by the way?
• 15:42
  Speaker 0: That is very cool.
• 15:44
  Speaker 0: It's not nice!
• 15:45
  Speaker 0: All right.
• 15:48
  Speaker 1: so I'm trying to open a can kind of works but then yeah pushing it in doesn't really work.
• 15:56
  Speaker 1: So apparently that's also something still good.
• 16:00
  Speaker 0: Oh its horrible.
• 16:01
  Speaker 1: Okay let me relieve you from that.
• 16:04
  Speaker 1: have a task for You But we need close your
• 16:07
  Speaker 0: eyes.
• 16:08
  Speaker 1: I put up couple of Euro coins here And you need to guess with closed eyes which ones are here, lying on the table right in front of your hands.
• 16:21
  Speaker 1: So there's four coins.
• 16:22
  Speaker 1: maybe can differentiate them?
• 16:24
  Speaker 1: Okay
• 16:25
  Speaker 0: I think it has been really close two minutes now and i've set a timer for my iPad.
• 16:31
  Speaker 0: so usually do ice baths but most horrible thing was always my toes.
• 16:37
  Speaker 0: So they're
• 16:42
  Speaker 1: here?
• 16:44
  Speaker 1: Okay,
• 16:46
  Speaker 0: I can actually feel... I can hardly grab them.
• 16:51
  Speaker 0: This one is small
• 16:53
  Speaker 1: so you can differentiate the size
• 16:57
  Speaker 0: and try to grab on.
• 17:01
  Speaker 1: No
• 17:01
  Speaker 0: this isn't possible right now?
• 17:02
  Speaker 0: really no it's not possible.
• 17:04
  Speaker 0: that was
• 17:04
  Speaker 1: great.
• 17:05
  Speaker 0: okay this one a small but i can only because obviously see how far there are
• 17:12
  Speaker 1: two more.
• 17:13
  Speaker 1: Can you differentiate those two?
• 17:15
  Speaker 0: Okay, let me see which one.
• 17:17
  Speaker 1: Those to yeah okay so So you differentiated by the size
• 17:24
  Speaker 0: only by the sides.
• 17:25
  Speaker 0: I can feel the texture now a little bit
• 17:28
  Speaker 1: coming back.
• 17:28
  Speaker 1: probably
• 17:30
  Speaker 0: it's getting more.
• 17:30
  Speaker 0: I have to put them back in
• 17:33
  Speaker 1: all right Yeah but i think I mean, everybody knows that from maybe winter.
• 17:38
  Speaker 1: If you are in a region where there is winter right?
• 17:41
  Speaker 1: But if he goes skiing for example... Everybody has had this thing that you have to deal with the zipper and
• 17:47
  Speaker 0: your
• 17:47
  Speaker 1: hands are cold.
• 17:49
  Speaker 1: so imagine now we have zero sense of touch at all!
• 17:53
  Speaker 0: So it was when i took them out first.
• 17:55
  Speaker 0: they were really zero-sense of such.. When i was trying to find the coins ,i thought i was touching the coins already but i wasn't.
• 18:02
  Speaker 0: because its like the sense is zero, right?
• 18:05
  Speaker 0: I feel like now i don't feel the difference between the table or the coins.
• 18:11
  Speaker 0: But then obviously touch comes back pretty fast but it's very hard to grab something.
• 18:16
  Speaker 0: and oh...I don't want to be a robot hand Yeah!
• 18:20
  Speaker 1: I mean this skin is the largest organ that human body has to offer.
• 18:25
  Speaker 1: It's just amazing
• 18:27
  Speaker 1: Right?!
• 18:27
  Speaker 1: It's so sophisticated And a robot.
• 18:32
  Speaker 1: even Fancy robots out there typically don't have that.
• 18:37
  Speaker 1: Of course, you can build You put in touch sensors.
• 18:41
  Speaker 1: They have to be calibrated.
• 18:43
  Speaker 1: they Can of course break because they're always getting into contact with things and then also only work in a certain range of force.
• 18:53
  Speaker 1: Suddenly suddenly you exceed that.
• 18:56
  Speaker 1: so if you want to start your AI based robot project today I'd rather look for tasks where you don't need that kind of thing.
• 19:05
  Speaker 1: Where things are working because they're relatively robust or... ...where you can see from a camera how things bend, change and then replaces the sense-of touch.
• 19:22
  Speaker 1: so if had done this with open eyes tons of world knowledge, you would be able to open the can because we know how to open a cam.
• 19:30
  Speaker 1: Because I've done it a million times
• 19:32
  Speaker 0: exactly so but as a gripper right?
• 19:34
  Speaker 0: The can wouldn't be a problem because i could grab.
• 19:36
  Speaker 0: that's not a problem cause when I have...I'm the robot.
• 19:39
  Speaker 0: now my hand is not frozen anymore as much.
• 19:42
  Speaker 0: So I could set this in the right tightness and always same And This Would Not Cause An Issue Right
• 19:50
  Speaker 1: Yeah But You Also Look At It As A Robot.
• 19:55
  Speaker 1: So then, you have seen it before.
• 19:58
  Speaker 1: You can see the can move and the robot has several cameras.
• 20:03
  Speaker 1: so... We use the camera as a replacement for touch.
• 20:09
  Speaker 1: in some respect?
• 20:10
  Speaker 0: It makes sense of course.
• 20:11
  Speaker 0: yeah.
• 20:12
  Speaker 0: I mean we do that.
• 20:15
  Speaker 0: sometimes we're used to something because now moves
• 20:21
  Speaker 1: Yeah.
• 20:21
  Speaker 1: I mean, the human has so much redundancy you can if you have ever gone to the bathroom at night?
• 20:28
  Speaker 1: You know that maybe don't even have to open your eyes.
• 20:31
  Speaker 1: It's
• 20:31
  Speaker 0: very true is very true.
• 20:33
  Speaker 0: um.
• 20:35
  Speaker 0: So tell me a bit more about Um The ai and where i adds value because I really want to understand like ai is doing so much right now And did.
• 20:46
  Speaker 0: the robots are obviously using more and more AI but Where, where does it make sense to actually use AI right now?
• 20:55
  Speaker 0: And why doesn't makes sense to us classical automation?
• 20:57
  Speaker 0: and maybe we're mixing things up too much.
• 20:59
  Speaker 0: So a few months ago there was this time when everybody is using AI for everything but then I'm Maybe that's not the correct approach.
• 21:07
  Speaker 1: Yeah For sure i think if you If You Use AI based robotics in cases where you could use program robot Mm-hmm Not set Up for Success because The Program Robot Is is optimized for that task.
• 21:22
  Speaker 1: It's better than a human doing their tasks, but then it has its limits and so every task we're doing... For example if you had one case where the company has hundred twenty different kinds of products which are all kind-of similar or they're sold in different quantities okay?
• 21:43
  Speaker 1: So there high runners have complete automation very, very competent company in terms of how the factory looks like.
• 21:53
  Speaker 1: It's an amazing piece of technology.
• 21:56
  Speaker 1: but then for one hundred and fifteen of those products they would still do some part manually because it wouldn't make sense to automate all that with all the effort that their using.
• 22:07
  Speaker 1: so now okay we're at a boundary where I don't want put all machinery.
• 22:17
  Speaker 1: I'm actually fine with, you know that thing maybe being slightly slower.
• 22:22
  Speaker 1: But i can deal With all these different weird situations That would otherwise have to model or remove from the whole process.
• 22:30
  Speaker 0: And thats where AI will solve this issue right?
• 22:32
  Speaker 0: Where it's ever changing.
• 22:34
  Speaker 0: Or is then the classical automation?
• 22:37
  Speaker 0: No its AI Right?
• 22:38
  Speaker 1: Yeah yeah.
• 22:39
  Speaker 1: so we're expanding The space into what We call semi-structured and Then later less and less structured environments
• 22:45
  Speaker 0: Okay.
• 22:46
  Speaker 0: It makes absolute sense, so obviously when you're doing something super fast highly automated it makes absolutely sense but then as soon as gets more I would say like unstructured less quantities than make sense to go and approach where can react on the environment?
• 23:01
  Speaker 1: yeah
• 23:01
  Speaker 0: very interesting.
• 23:03
  Speaker 0: So how do we get through that process of deciding?
• 23:07
  Speaker 0: What is a useful mental model for this?
• 23:09
  Speaker 0: To understand what actually could be done in your office Or you're sitting, looking at the factory floor and thinking about how where could I implement a robot?
• 23:20
  Speaker 0: A new robot technology like that from Robco right now.
• 23:23
  Speaker 1: Right so... So i think one question to ask is can i break down the task into atomic units of some form?
• 23:32
  Speaker 1: And um..and How complex Is The Whole Thing?
• 23:37
  Speaker 1: We are talking About robots That Are Learning From Examples And this works best.
• 23:44
  Speaker 1: if I think about a very easy like mental model that an AI model, Think of it as memory off movements.
• 23:59
  Speaker 1: Right?
• 24:00
  Speaker 1: So i'm showing at movements memorizes them.
• 24:04
  Speaker 1: That's obviously not what you would want an eye system for.
• 24:09
  Speaker 1: You want to have to generalize right.
• 24:11
  Speaker 1: so yeah does more than.
• 24:13
  Speaker 1: But that's, the easiest thing to think about.
• 24:16
  Speaker 1: So now there is like this additional factor generalization where all the wisdom from a million different cases have already gone into this model.
• 24:29
  Speaker 1: The model gets bigger and bigger and so on.
• 24:32
  Speaker 1: There's basically an additional thing you get from foundation models world models what we talked.
• 24:41
  Speaker 1: That factor is getting bigger every year, every quarter actually.
• 24:46
  Speaker 1: But I think it still makes sense to have this mental model and then okay so now you have movements that need to make decisions in between.
• 24:55
  Speaker 1: So at the decision point there's basically a new probability distribution And those multiply.
• 25:00
  Speaker 1: You get combinatorial explosion.
• 25:03
  Speaker 1: Then you probably need somehow think about how do we split them up?
• 25:07
  Speaker 1: Do i give language constraints, for example.
• 25:11
  Speaker 1: So suddenly it can make sense of what you're doing here where system there.
• 25:15
  Speaker 1: so these things are done.
• 25:19
  Speaker 1: but obviously if your first AI project have something simple not with too much variation easy to implement and also limited amount of data that you need Data becomes Less, less data is needed over time.
• 25:42
  Speaker 1: But right now this is where we're at Right?
• 25:49
  Speaker 0: Yeah I find that the data part.
• 25:51
  Speaker 0: it's also interesting because i think in the beginning when you are looking for a project Usually You have to collect your data first and That can take some time.
• 26:00
  Speaker 0: So taking the first step into robotics In new era Makes sense as a small project, because then you have the success.
• 26:11
  Speaker 0: And then you can much better imagine what the next step could look like and I think planning it too big and being disappointed that doesn't work in sync on huge projects also lead to problems where not everything is solvable but stuck with some parts of between.
• 26:33
  Speaker 1: You always want start smaller these projects.
• 26:35
  Speaker 1: so one thing for example being able to differentiate a lot of different products, SKUs.
• 26:44
  Speaker 1: It is for example pretty easy to build the robot that can deal with packages that come from a conveyor belt and then you need to split in two parts like two outputs for example.
• 26:57
  Speaker 1: because what it needs to learn is what as package?
• 27:00
  Speaker 1: And basically be able to treat every package in sort of same way.
• 27:04
  Speaker 1: yeah What's much harder if If you are, let's say a supermarket and you have eight thousand different products.
• 27:11
  Speaker 1: And do want to differentiate all eight thousand of them?
• 27:14
  Speaker 1: Yeah because I mean if he did it naively would mean just collecting whole bunch of data for every product so that doesn't scale.
• 27:23
  Speaker 1: So then you have to think about can i do hybrid approach where there is scan in between or Can I group products into the things that interest me, for example?
• 27:40
  Speaker 1: Into boxed products.
• 27:42
  Speaker 1: And so all of them are the same and they just need to be handled in the same way.
• 27:50
  Speaker 1: So these other kind of simplifications we're doing
• 27:53
  Speaker 0: When you looking at infrastructure like compute power network power as well.
• 27:59
  Speaker 0: Like obviously robot is standing there.
• 28:02
  Speaker 0: How's it usually connected network.
• 28:06
  Speaker 1: Does
• 28:06
  Speaker 0: it have a data cable, Wi-Fi probably?
• 28:09
  Speaker 0: And then how much computing happens inside the robot and how much competing happens outside of the
• 28:14
  Speaker 0: robot?".
• 28:14
  Speaker 1: Yeah so obviously we want to do as much possible on the robot itself.
• 28:22
  Speaker 1: We want... I mean there's all the low level stuff which is you know There.
• 28:26
  Speaker 1: anyway in that most compute heavy part are these intelligent models and imagine You know industry ready.
• 28:37
  Speaker 1: And so the best compute that also works on an AMR when I'm moving around, um has something like half of a compute power from our gaming workstation.
• 28:48
  Speaker 0: Oh really?
• 28:48
  Speaker 0: Um
• 28:50
  Speaker 1: So-so That's-that's the limit.
• 28:51
  Speaker 1: it has you know quite a bunch of memory.
• 28:53
  Speaker 1: It's also unified memories or in that sense is already pretty good machine.
• 28:59
  Speaker 1: but its Like doesn't compare to a cluster that you use when you run a query against chat GPT, for example.
• 29:11
  Speaker 1: Then we're talking about a cluster, seventy-two GPUs maybe five to ten million dollars and then it only responds if your tokens per second right?
• 29:23
  Speaker 1: So so... That's the limit!
• 29:26
  Speaker 1: We can't just say hey let's take this language model at me using just apply two robotics.
• 29:32
  Speaker 1: people are doing this now But it's like a thousand times too slow, right?
• 29:37
  Speaker 1: And then there is the middle ground.
• 29:38
  Speaker 1: You can say okay so Tony takes five milliseconds to do that cloud transfer.
• 29:44
  Speaker 1: but if you now I do part on the robot and apart from let us say one beefy GPU machine in the Cloud um...you pay for with less guarantees.
• 29:55
  Speaker 1: There are lot of jitter going on.
• 29:57
  Speaker 1: Um..You don't really understand anymore where this is coming in, there's obviously the whole safety aspect and security aspect around it.
• 30:07
  Speaker 1: So we're trying to do as much possible onsite at reasonable cost.
• 30:12
  Speaker 1: And that limits right now... This gives a hard upper limit on the compute you can actually afford which also means intelligence.
• 30:21
  Speaker 1: today That was something better than two years ago.
• 30:29
  Speaker 1: Also, I mean if you see the developments computers running matrix multiplications which are at the heart of neural nets Have become maybe ten to a hundred million times faster than in the mid-eighties when this was first started To being used.
• 30:47
  Speaker 1: so there is more On the way.
• 30:50
  Speaker 1: I'm sure it's gonna be ten Maybe a hundred times faster.
• 30:53
  Speaker 1: we look down the road now that the bottleneck is mostly memory transfer.
• 31:02
  Speaker 1: On the other hand, you know these chips are just ten-twenty atoms in size every transistor.
• 31:09
  Speaker 1: So there's some physics involved here which make it harder and harder to improve that.
• 31:15
  Speaker 1: I'm sure we'll see photonic chips Which are kind of analog computers coming In.
• 31:21
  Speaker 1: but You Know There is a tonne Of Things Coming.
• 31:23
  Speaker 1: But Physics Is A Heart Limit
• 31:26
  Speaker 0: Right?
• 31:26
  Speaker 1: So, so we try to make do with what we have right now.
• 31:31
  Speaker 1: And and so they're therefore probably not as far as you know modeling a complete human brain like this.
• 31:39
  Speaker 0: Yeah I mean obviously that's why i thought Like That's Probably What The Expectation Is.
• 31:44
  Speaker 0: As Well Maybe In Some Of The Clients When They Are Asking You Know JGPT Or Entropic LLMs Complex Questions And They Get Amazing Answers.
• 31:55
  Speaker 0: And I think people are transporting this idea that the experience they're having on their computers and devices where a huge data center processes, the request.
• 32:04
  Speaker 0: Then gives it back for very little money to be honest but obviously like maybe that is very subsidized.
• 32:09
  Speaker 0: But you know there's a huge computer working right?
• 32:12
  Speaker 0: People will say okay Can't be that hard to give a task to a robot?
• 32:18
  Speaker 0: then does that Obviously what i think many people fail understand?
• 32:23
  Speaker 0: The compute power used by the robot just to control its arms, to hold it's balance.
• 32:27
  Speaker 0: To be in this physical space is already quite a lot I'm pretty sure.
• 32:32
  Speaker 0: and then obviously you can't react as fast.
• 32:37
  Speaker 0: so that makes lots of sense.
• 32:38
  Speaker 0: It also make sense That You Can't Give Too Complicated Commands And too much data at the same time because the processing doesn't work
• 32:46
  Speaker 1: Sure!
• 32:46
  Speaker 1: Also the energy usage right?
• 32:48
  Speaker 1: Such a cluster than talked about has maybe hundred kilowatts
• 32:51
  Speaker 0: True
• 32:52
  Speaker 1: And it serves similar to a main frame.
• 32:54
  Speaker 1: It serves maybe a couple of hundred users at the same time, otherwise there wouldn't be an economic path there.
• 33:01
  Speaker 1: but then robot needs to be just online with very tight guarantees all the time.
• 33:06
  Speaker 1: so its'a very different setup
• 33:08
  Speaker 0: and obviously safety is also concerned right?
• 33:10
  Speaker 0: So like you can't... With these LLMs they hallucinate.
• 33:15
  Speaker 0: The more complex or open-the-world model this the more corner cases happen that where something could go different than expected.
• 33:23
  Speaker 0: And obviously with a tighter frame, I think it makes it safer and...
• 33:29
  Speaker 1: Right.
• 33:30
  Speaker 1: so talking about safety the idea when you see human robot can react like a human from a safety standpoint.
• 33:44
  Speaker 1: that requires of course all sorts of sensors.
• 33:47
  Speaker 1: if we put them in these standards On this level of safety, four humanoids are still evolving.
• 33:55
  Speaker 1: I haven't really been written like signed off.
• 34:02
  Speaker 1: so that means if you do your first project just assume industrial safety standards.
• 34:08
  Speaker 1: obviously our robots implement those safety standards but it's basically he.
• 34:12
  Speaker 1: stop fences maybe laser fences.
• 34:16
  Speaker 1: Just keep it simple
• 34:17
  Speaker 0: Because the robot is not a human.
• 34:24
  Speaker 0: If you bump into a robot, You'll know that it's Not A Human!
• 34:26
  Speaker 0: I have this great example where we were at an affair...I told you this already Where there was a robot dog and i was jumping around And everybody loved The Robot Dog.
• 34:33
  Speaker 0: It maybe two or three years ago.
• 34:34
  Speaker 0: This Is Everybody Seen The Dog.
• 34:36
  Speaker 0: Its Actually Pretty Aggressive But It Looks Cute.
• 34:40
  Speaker 0: One guy jumped basically like next to it Into the sub-movement and the doctor stopped and rammed in To him And I heard him quite a lot because it was pure metal hitting your flesh and bones.
• 34:52
  Speaker 1: Yeah, so you gotta be careful around that.
• 34:56
  Speaker 0: Is there already cases where robots in humans are working like basically next to each other?
• 35:01
  Speaker 0: In certain safe spaces or cases?
• 35:04
  Speaker 1: Right!
• 35:04
  Speaker 1: So this whole notion is called cobot... That's something that was developed ten-fifteen years ago ...and the idea is that you have some pretty reactive Safety that would slow down a few if you get in kind of the danger zone and And it's not used.
• 35:24
  Speaker 1: A lot hasn't really Hasn't really become very popular.
• 35:29
  Speaker 1: I think something like ten fifteen percent more robots because In the end, You pay The price with Slowdown
• 35:36
  Speaker 0: okay?
• 35:37
  Speaker 1: So If you know if you run Production line you just optimize for one thing right then so typically you end up going back to keep the robot space safe and inside, it's not safe.
• 35:52
  Speaker 1: Then you can run at full
• 35:54
  Speaker 1: speed.".
• 35:57
  Speaker 1: The next step would be really kind of stuff that is implemented in autonomous driving cars where its all about prediction because predictions need AI models as well.
• 36:11
  Speaker 1: And so when they develop those code standards, you know that wasn't really a topic.
• 36:16
  Speaker 1: Um That had in mind.
• 36:18
  Speaker 1: So prediction means You have enough sensors so that you can look one second two seconds into future Mm-hmm.
• 36:25
  Speaker 1: and so now because you have to break for example You Can start breaking Because you think something is will go In your way?
• 36:32
  Speaker 1: I think this.
• 36:33
  Speaker 1: This Is where we're going but not Something that he would do in Your first installation right Now.
• 36:40
  Speaker 0: So how do I have to think about this?
• 36:42
  Speaker 0: If say, okay.
• 36:45
  Speaker 0: I've looked at the process we're having in our factory production cycle and figure something out where we can deploy a robot that makes sense it's safe.
• 36:56
  Speaker 0: It simple is not too crazy.
• 36:58
  Speaker 0: We want get started with this.
• 36:59
  Speaker 1: then
• 37:00
  Speaker 0: How does that do you like?
• 37:01
  Speaker 0: order robots set up on its go time or how does implementation work
• 37:06
  Speaker 1: right?
• 37:07
  Speaker 1: so i mean there are obviously cases.
• 37:12
  Speaker 1: They're already known.
• 37:14
  Speaker 1: Simple things that we've shown today, which are kind of copy-paste and they call those jobs.
• 37:21
  Speaker 1: so those are relatively simple to set up.
• 37:25
  Speaker 1: And then right now there's still a ton of cases where We look for collaborations partnerships Where we setup what do you call it program?
• 37:35
  Speaker 1: So it's kind of a joint collaborative project where we go through, you know different implementation stages.
• 37:42
  Speaker 1: Mm-hmm.
• 37:44
  Speaker 0: so that... That's very interesting.
• 37:45
  Speaker 0: do have like uh um A list of things people should think about when Like a checklist for sequencing for decision makers?
• 37:56
  Speaker 1: I Think in this case whether It is a Java program.
• 37:58
  Speaker 1: You mean i think This really on our case by case basis because the experience On our side obviously also growing every month.
• 38:05
  Speaker 1: And so, yeah one would have to talk with you know.
• 38:08
  Speaker 1: But
• 38:08
  Speaker 0: you can help of that right?
• 38:09
  Speaker 0: You have the jobs installed already on this software or all the programs and everything Right.
• 38:15
  Speaker 1: I think we're at a stage where We look different projects individually then need to make decision.
• 38:21
  Speaker 1: i think it will become simpler every couple months but thats very standard new technology.
• 38:31
  Speaker 1: And yeah, there's.
• 38:32
  Speaker 1: you can't just buy it at Amazon and.
• 38:36
  Speaker 1: Just start doing it.
• 38:37
  Speaker 1: for if we talk about vision that's where we're going?
• 38:40
  Speaker 0: Yeah I think is already become a huge step.
• 38:44
  Speaker 0: because thing about this as we talked about like smaller or simpler versions of production steps where you can implement the robot.
• 38:54
  Speaker 0: This was not possible many years ago Because yet to think about the whole production line making it possible to have individual robots, like taking over new parts of the production process is already very special because they are specifically catered for a certain part and this was not possible before.
• 39:15
  Speaker 0: Everything had to work together in sync and software was complicated.
• 39:19
  Speaker 0: now that's changing a lot.
• 39:23
  Speaker 1: I think everything was just set up by hand programmed IoT, you know big systems.
• 39:31
  Speaker 1: I think it just becomes more and more natural.
• 39:33
  Speaker 1: It will use natural language to communicate the wheel in the end be a fleet that You can have a look and Observe and control.
• 39:42
  Speaker 1: but you know what?
• 39:43
  Speaker 1: Its its different era
• 39:45
  Speaker 0: Absolutely That maybe one last question to this.
• 39:47
  Speaker 0: And there's also like if you you set up The robot and then you figure out Like through two months later when you figure Oh we have A better way of doing This Then how'd you go about that?
• 40:00
  Speaker 0: Or maybe say, okay.
• 40:01
  Speaker 0: No I have an even better idea to use this robot or Maybe add some more robots.
• 40:05
  Speaker 0: How do you go bout the change.
• 40:08
  Speaker 1: Yeah i mean we're in a very data oriented regime.
• 40:12
  Speaker 1: of course This is also one big difference.
• 40:16
  Speaker 1: First of all when you- Even When You start setting it up We always Say The Robot Has To Be Connected Right.
• 40:23
  Speaker 1: So you can like an earlier time say, no I don't want any connection to the outside world.
• 40:29
  Speaker 1: that is really would be really hard to do.
• 40:34
  Speaker 1: just from data volume it's at least a hundred times more data volume than we had before talking about roughly gigabyte per minute of raw data.
• 40:45
  Speaker 1: so if you wanna just collect data obviously compress it right?
• 40:52
  Speaker 1: You can compress it, you need to move into the cloud where training happens because GPUs are not powerful enough on site.
• 41:00
  Speaker 1: We're saying we want at least a hundred megabit connection Not a GSM connection or an LTE connection.
• 41:10
  Speaker 1: So no Wi-Fi connection either in best case unless its movable put a big pipe there.
• 41:20
  Speaker 1: Obviously, there's safety and privacy things that we're taking care of off course.
• 41:29
  Speaker 1: this is all part of the whole process.
• 41:32
  Speaker 1: how to set it up as something would be discussed but in our space where its similar to lets say Tesla or one these other car companies which collect data and then on a case-by-case basis use that to improve the products over time.
• 41:53
  Speaker 1: And I think, you know it's network effecting.
• 41:55
  Speaker 1: everybody has benefits from doing that.
• 41:59
  Speaker 1: there is also possibility of remote maintenance.
• 42:05
  Speaker 1: Also something needs be safe but can take care off.
• 42:10
  Speaker 1: Talking with companies this always something where they say This is our IT department.
• 42:17
  Speaker 1: I need to, you know have a six months before i can get this done.
• 42:23
  Speaker 1: so just be aware that it's something required and different from how we've been doing things.
• 42:41
  Speaker 0: Maybe like as a last information, what would you say to somebody who's thinking about it right now?
• 42:47
  Speaker 0: To implement AI robots and AI in their production.
• 42:53
  Speaker 0: What is the things that they should watch out for right now just as a summary of this podcast how they should move forward?
• 43:03
  Speaker 1: Yeah so I think you want to be set up for success.
• 43:07
  Speaker 1: first thing is don't assume job being taken over by a humanoid that has the same properties.
• 43:16
  Speaker 1: Think simple thing as an extension, think about cases which are not possible to do in the rigid classical robotic space?
• 43:30
  Speaker 1: I think what we said with grippers.
• 43:32
  Speaker 1: can you do it?
• 43:40
  Speaker 1: Do I have the connectivity?
• 43:42
  Speaker 1: And do I have a right working relationship with vendor, and join project that is set up for just showing value we can implement.
• 43:53
  Speaker 1: That already gives us a ton of possibilities.
• 43:59
  Speaker 1: even you know all these restrictions put in.
• 44:02
  Speaker 1: there's so much work to be
• 44:08
  Speaker 0: done robotics, obviously it's like or robotics with AI.
• 44:13
  Speaker 0: Or this new physical AI is a movement that's happening right now and do you see it as something where You should wait on moving on?
• 44:22
  Speaker 0: It always something that
• 44:24
  Speaker 1: you should
• 44:25
  Speaker 0: look at right now And figure out how its working.
• 44:27
  Speaker 0: so they are set up for success in the future.
• 44:30
  Speaker 1: Well I think different people have different preferences there.
• 44:34
  Speaker 1: either The company just waiting isn't a market when They can Wait.
• 44:39
  Speaker 1: Others have very concrete needs.
• 44:43
  Speaker 1: We have in Germany this whole wave of retirements that is coming up the next five years, so people want to just make sure they can run their business for a couple year and then try things out start moving learn from it quickly I hope what we talked about today Because I think what would be really bad is if you start something with super high expectations and then it falls flat.
• 45:14
  Speaker 1: And, uh... You know?
• 45:17
  Speaker 1: People started looking at that five years later because your competitors in China or other countries might have overtaken you.
• 45:26
  Speaker 0: Yeah!
• 45:27
  Speaker 0: It didn't work once setting up the runway.
• 45:29
  Speaker 0: so one of the important things also If you're thinking about a project and want to set-up for success you should definitely talk to the team at RobCo.
• 45:39
  Speaker 0: And, I think that's always open-to-talk and just give them information they need to understand if this is possible or not?
• 45:49
  Speaker 0: Is it right solution for
• 45:50
  Speaker 0: them?".
• 45:51
  Speaker 1: Yeah!
• 45:52
  Speaker 0: That's a recommendation because i know how complicated stuff is.
• 45:57
  Speaker 0: It makes sense... It isn't about making success of one robot but Attaching to the future of what's happening and not saying, oh no this doesn't work.
• 46:07
  Speaker 0: I'm going look at it again in ten years And then ten years are buying.
• 46:09
  Speaker 0: everything has changed.
• 46:11
  Speaker 0: so i think that's great and i think you guys Are.
• 46:13
  Speaker 0: um have such capable team members who love To talk to anyone out there?
• 46:18
  Speaker 0: Who wants to set up the robots?
• 46:20
  Speaker 1: Yeah, man It's a learning experience for everyone.
• 46:23
  Speaker 1: i think we Have made A lot Of steps around it and yeah it's good to start on The same level.
• 46:30
  Speaker 1: So That We have You know, we can move faster.
• 46:34
  Speaker 1: If the expectations are set correctly and if somebody comes in with something that is very realistic anywhere in the first place.
• 46:42
  Speaker 0: Well thank you Clemens.
• 46:43
  Speaker 0: this was amazing.
• 46:44
  Speaker 0: I had so much fun cold hand as well for the ice bath.
• 46:47
  Speaker 0: Amazing.
• 46:48
  Speaker 0: And um i think our viewers got a chance to understand a bit more how This works today?
• 46:53
  Speaker 0: What's possible Today into just have their expectations right
• 46:56
  Speaker 1: To
• 46:57
  Speaker 0: make The project A success.
• 46:58
  Speaker 0: So Thank You Clemens.
• 47:00
  Speaker 0: We'll talk soon again when we're back with a new episode of Rob Talk.
• 47:04
  Speaker 0: And if you like the episode, obviously have to subscribe.
• 47:08
  Speaker 0: share it your friends and colleagues who are also in this space?
• 47:10
  Speaker 0: Maybe somebody is deciding on using robots or maybe who's got production floor where he isn't using robots yet in an adequate amount.
• 47:18
  Speaker 0: send that them then see us at RobCo and talk more!
• 47:30
  Speaker 1: Subscribe on Spotify, Apple Podcasts and YouTube.
• 47:33
  Speaker 1: See you on the factory floor!