Update der Roboterhand von Shadow Robot Company

00:00:00: Hi, Robert hier.

00:00:01: Dieser Podcast wird euch präsentiert von der Hannover Messe.

00:00:04: Wir sagen vielen herzlichen Dank an die Kollegen in Hannover und jetzt geht's los.

00:00:09: Robotik in der Industrie, der Podcast mit Helmut Schmidt und Robert Weber.

00:00:18: Hi, Robert hier.

00:00:22: Das ist die letzte Folge vor unseren Sommerferien.

00:00:27: Und wir haben uns dieses Mal bei den Kollegen vom Industrial AI Podcast

00:00:31: bedient, nämlich die hatten Rich schon vor einigen Wochen in ihrem Programm.

00:00:36: Und der kam so gut an, dass wir gesagt haben, OK, nehmen wir Rich Walker von der

00:00:40: Shadow Robot Company auch mit in den Robotik in der Industrie Podcast.

00:00:46: So Helmut und ich gehen jetzt in den Urlaub, entspannen uns ein bisschen.

00:00:51: Der Helmut bleibt noch ein bisschen länger.

00:00:52: Der fährt nämlich erst Anfang September in den Urlaub.

00:00:55: Ich fahre jetzt in den Urlaub und wenn wir wieder da sind, gibt's was ganz

00:01:01: Besonderes, was Neues vom Podcast Robotik Industrie.

00:01:05: Wir wechseln nämlich die Seiten.

00:01:07: Was das bedeutet, erfahrt ihr dann in der

00:01:09: erste Folge nach den Ferien.

00:01:11: Wir wünschen euch alles Gute und bleibt vor allem gesund.

00:01:14: Und jetzt geht's los.

00:01:17: Mein Gast heute ist Rich Walker.

00:01:19: Er arbeitet für die Shadow Robot Company.

00:01:21: Und Rich, das ist dein zweites Mal als Gast auf unserem Podcast.

00:01:25: Willkommen zurück.

00:01:26: Danke, es ist immer ein pleasure to be here.

00:01:28: Lass mal, dein Kat, der auch eine

00:01:30: kruzische Part in der Podcast ist ein Kat auch da.

00:01:33: Er ist heute zu Hause und ich bin in der

00:01:35: Office, so wir hoffentlich werden die

00:01:37: Feline Intervention sparen.

00:01:38: Okay, wie geht es dir?

00:01:41: Sehr gut, sehr gut.

00:01:42: Wir haben einfach die neue Robot Hand finally

00:01:43: released, die neue Robot Hand, die wir in

00:01:46: Secret für fünf Jahre arbeiten.

00:01:47: So, das ist eine sehr tolle,

00:01:49: etwas expansiv Russische, sagen wir uns.

00:01:51: Und wir sind sehr froh, zu sagen, dass wir

00:01:52: Leute darüber sprechen können.

00:01:53: Genau.

00:01:54: Ich rede über AI und Robotik.

00:01:57: Nvidia ist mit den Händen mit Hardware.

00:02:00: Und so bist du.

00:02:01: So ist Hardware noch wert?

00:02:04: Oh, absolut.

00:02:05: Ich denke, was wir mit den letzten fünf

00:02:07: Jahren gesehen haben, ist, dass es in der

00:02:08: AI-Action total drifft ist, mit viel, viel besserer Comput.

00:02:13: Nvidia ist ein sehr großer Beton für GPUs und Fabriken, für die deepen

00:02:17: Erleichterung. Das hat natürlich wirklich, wirklich

00:02:19: geklappt für sie und für jemanden, der einen Video-Scheinungen

00:02:21: hat, fünf Jahre ago, gut gemacht.

00:02:24: Für die restlichen.

00:02:26: Hast du Schäden?

00:02:27: Ich habe nicht in Nvidia.

00:02:28: Fortschließlich, da gehen wir.

00:02:30: Ich bin mein eigenes Haufen.

00:02:34: Aber für uns über die Hardware, das ist ein deep, deep-tech Corner,

00:02:38: ich denke, das ist etwas, das wir sehen, dass du nicht neue

00:02:41: Hardware haben, ohne viel Arbeit, viel Entwicklung.

00:02:43: Aber du musst neue Hardware haben, um alles möglich zu machen.

00:02:47: Wie viel Software und AI ist in deiner neuen Hand?

00:02:52: Remarkabel little.

00:02:54: Wir hatten eine schöne Relationschaft, mit Google Deed Mine,

00:02:57: entwickeln diesen Robots.

00:02:59: Und die haben gesagt, wir sind die AI-Präsidenten,

00:03:01: du bist die Roboter-Hardware-Präsidentin.

00:03:03: Hier ist eine kleine Liste.

00:03:04: Lass uns nicht zu viel über die Liste stecken.

00:03:07: Und das war eigentlich sehr, sehr rewarding,

00:03:09: weil es dachten, wir mussten nicht mit den

00:03:11: Herausforderungen, die sie mitnehmen.

00:03:12: Aber auch können wir sie sagen, wir haben

00:03:15: ein paar Ideen über die Hardware-Präsidentin.

00:03:17: Wollen sie das Problem lösen, die sie haben?

00:03:19: Und sie sagen, ja, das wird funktionieren.

00:03:21: Und das wird funktionieren.

00:03:22: Lass uns beide tun.

00:03:24: Was ist jetzt neu?

00:03:25: Was ist in AI und Robotik im Moment geändert?

00:03:29: Die Veränderung, die wir gesehen haben, ist diese

00:03:32: Algorithmen aus dem Video-Game und in der realen Welt.

00:03:37: Wir haben viele, viele Examples gesehen, wo

00:03:39: wirklich powerful AI-Systems haben

00:03:41: und mit den complexen Taschen zu tun.

00:03:43: Aber die sind fast all die virtuellen Taschen.

00:03:47: Wenn man sie auch in der realen Welt tun könnte,

00:03:49: dann waren sie virtual.

00:03:50: So, wenn man sich auch auf Alpha-Folge,

00:03:52: die Protein-Folge, schafft,

00:03:54: aber es ist nicht das Experte zu verwerfen,

00:03:57: die Protein-Folge, das wird noch by Chemisten gemacht.

00:04:01: Und was wir versuchen, über einen anderen

00:04:03: Korn der AI-Spaces zu machen, ist, dass wir sagen,

00:04:05: OK, wir nehmen diese powerful AI-Tools und

00:04:08: wir nutzen sie auf realen Robots, um die Probleme zu lösen,

00:04:12: die für einen langen Zeit in Robotik im Stuck sind.

00:04:15: Zum Beispiel?

00:04:17: So, wie picken wir Objekte und interaktieren sie gut?

00:04:20: Das ist ein wirklich, wirklich großer Herausforderung in Robotik.

00:04:22: Wenn man sich alle sehr

00:04:24: erfolgreiche Unternehmen, die mit Robots

00:04:25: Handeln tun, schaut man, was sie tun,

00:04:28: sie sind sehr, sehr limitiert in den Typen,

00:04:30: die sie handeln oder die Environments, die sie tun.

00:04:33: Auch die Weltklasse-Experten, die

00:04:36: Handeln Objekte auf der Stelle, wie Amazon,

00:04:38: sind noch sehr, sehr limitiert in den Domain.

00:04:41: Amazon erwartet das Objekt, das auf der Strecke auf der Strecke.

00:04:44: So wie ein Kardo, weiß, dass das Objekt in einem Warehaus wird.

00:04:49: Aber wenn du dich anstatt, hier ist ein randomes Teil

00:04:51: des Sachen in deinem Kinder-Bedroom, wie ich sie umschen?

00:04:53: Ja. - All diese Systeme, einfach

00:04:55: throw die Hände in den Herd und sagen, ja,

00:04:57: du kannst noch nicht das tun.

00:04:58: Du hast alle diese AI-Auguristen gesagt.

00:05:02: Wie integriert du jetzt, wie du die bothen Wörter in

00:05:07: deinem Hardware-Wert und dein AI-Auguristen wohnen?

00:05:11: So, ein lot of that happens through a tool called Reinforcement Learning.

00:05:15: Reinforcement Learning is really powerful.

00:05:17: It's a method that allows you to take advantage of the fact that people have got

00:05:22: really good at simulating physics.

00:05:24: I think Festo already presented a hand with a Rubik, I think, was it?

00:05:29: Yeah, exactly.

00:05:30: I think quite a lot of interesting challenges where people have used

00:05:33: Reinforcement Learning to solve them.

00:05:34: But typically the problem you have is that,

00:05:37: okay, you can do that in the simulator and it works really, really well.

00:05:40: But then when you try and run those experiments in the real world,

00:05:44: what happens is you break your robot.

00:05:46: And trying to keep the robots running for long enough to get these experiments done

00:05:50: has been essentially the big showstopper here.

00:05:53: Yeah.

00:05:53: And what is now your approach?

00:05:57: Well, we spent thousands of hours beating the robots with sticks, so you didn't have to.

00:06:02: Is the honest simplest way to it, I can explain.

00:06:05: For the new hardware, we said, okay,

00:06:07: what Reinforcement Learning does is it tries random movements and works out,

00:06:12: which ones are good quality movements.

00:06:13: So we have to be able to survive that.

00:06:15: And then once we can survive that,

00:06:17: what grasping a manipulation does is it collides the robot hand with objects.

00:06:22: But if I'm going to pick something up,

00:06:24: my fingers are colliding with the object by definition.

00:06:27: They have to be to touch it.

00:06:28: And traditionally in robotics, we avoid collisions.

00:06:30: We say, plan your path so you don't collide.

00:06:34: But if you want to learn to handle objects well,

00:06:36: you have to accept that collisions are going to be your bread and butter.

00:06:39: Exactly.

00:06:40: We built a robot platform that was designed to survive

00:06:44: quite hostile abuse.

00:06:45: We hit it with pistons, we hit it with baseball bats.

00:06:48: We hammered it into itself and over thousands of hours of testing.

00:06:53: We ended up with a platform that was reliable enough to perform these

00:06:57: experiments that really make you wince when you see a robot running on them.

00:07:01: So you're running the hand together with the robot, collecting data?

00:07:06: Am I wrong?

00:07:07: Yes, exactly.

00:07:08: The learning algorithm runs partly in simulation,

00:07:11: but then translates across the real hardware.

00:07:13: When running on the real hardware,

00:07:14: we have as much data as we can possibly get coming out of the robot.

00:07:19: So this is something that people don't appreciate about classical robots.

00:07:22: They don't actually generate very much information.

00:07:25: You might have some motor current data and you might have some joint position

00:07:28: sensors.

00:07:29: Und if you're really lucky, you'll have a force talk sensor at the wrist.

00:07:32: And that's probably it.

00:07:34: As if you're trying to learn to understand acting in the real world,

00:07:38: you want as much data as possible.

00:07:41: So we've packed this robot with sensing.

00:07:43: You have really, really deep insights into what's

00:07:45: happening around the robot and with the robot, what it's working.

00:07:49: And how much data do you collect?

00:07:51: And what kind of algorithms now do you use then with the data?

00:07:56: So we have, I think it's something like a hundred and fifty five sensors

00:08:00: or each finger of the robot.

00:08:01: Each finger,

00:08:02: but you still have three fingers, right?

00:08:05: Three fingers on this, three fingers.

00:08:06: And then we have.

00:08:06: That's new, right?

00:08:07: It's our previous hands were always very humanlike.

00:08:10: So there'd be four fingers and a thumb and the fingers would be that kind

00:08:14: of characteristic up and down pattern that your fingers come in.

00:08:18: The thumb comes up to propose those.

00:08:20: This new hand is actually, it's only three fingers,

00:08:23: but each one of those fingers is probably as agile as your thumb is.

00:08:28: And there are a range of triangles so they could do quite a lot of manipulation,

00:08:31: quite a lot of dexterity.

00:08:32: We could put more fingers there,

00:08:34: but for the moment three was enough to do interesting things and see what could

00:08:38: be done and how we could do it.

00:08:39: And now you have this one hundred.

00:08:42: You mentioned one hundred fifty sensors, right?

00:08:44: Yeah. In one finger, right?

00:08:45: In one finger.

00:08:46: OK, now you collect the data and what is the next step then?

00:08:51: What do you do with the data?

00:08:53: So for somebody doing research in robotics in AI,

00:08:56: what they're typically doing with that data is they're taking it and they're

00:09:00: feeding it into the capturing it while they're doing an experiment and they're

00:09:03: feeding it into their learning system.

00:09:05: So that might be going in as real time data that's used to generate control.

00:09:09: You feed it into neural networks.

00:09:11: Neural network generates commands back to the robot or it might be something

00:09:15: where you're using a demonstration and then at the end of the process,

00:09:18: you're saying, OK, capture this data and learn from this data how to control the robot.

00:09:23: And tools like lots of kind of good learning tools out there in the in the

00:09:27: ecosystem now, PyTorch and so on, people just feeding this data into.

00:09:31: So, but this is not the reinforcement approach, right?

00:09:34: This is how the the hand behaves, right?

00:09:37: Well, the reinforcement approach is the next step, right?

00:09:40: The simulation stage, typically.

00:09:42: And then also on the real robot.

00:09:44: But what the reinforcement system has to learn to do is to use both the images

00:09:49: of robots in the world and the data from the robot to control its behavior.

00:09:55: And of course, something that's very interesting is when your simulation

00:09:59: and your real robot have different properties.

00:10:01: Sometimes it's called a sim to real gap because it's easier to do things in

00:10:05: simulation than reality, but on other occasions, we actually have more data

00:10:10: available in reality, because we can we can detect things that are actually

00:10:13: very, very hard to model in simulation and you're getting real live data from

00:10:17: the world, from the interaction.

00:10:19: How many real live data does your hand need to see or not to see to?

00:10:25: How do you call it?

00:10:27: Well, it's interesting if you apply,

00:10:29: we can go back to classical control approaches with this hardware.

00:10:32: So typically internally, when we organise a demo, we're using position

00:10:36: control or joint control, because we implement that on the robot for you.

00:10:42: We're not trying to do the high level learning systems,

00:10:45: but if we are, people who are, they have all this extra data.

00:10:48: So they have things like fingertip, Kontakt data, video streams from the fingers.

00:10:53: There are IMUs, so you can measure vibration systems.

00:10:56: And then there are kind of the low level information you expect,

00:10:58: like the motor temperature and the motor current and things.

00:11:01: Lots and lots of data so you can build a,

00:11:04: you've got the ability to build a wealth of insight,

00:11:07: whether your algorithm ends up using it or not, is up to you.

00:11:11: OK, OK.

00:11:12: Can you, because we already talked about your new approach,

00:11:16: could you please share some technical key facts?

00:11:19: Is it an update? Is it a total new design?

00:11:23: Rich, some updates.

00:11:25: We went right the way back to a clean slate for this, because we said to ourselves,

00:11:29: you know, we built anthropomorphic hands in the past, we've done lots and lots of those.

00:11:32: But actually maybe we don't care about anthropomorphism, we care about dexterity

00:11:37: and reliability, and that turned out to be a different kind of lens to look at the

00:11:42: problem for. So we were able to start and say,

00:11:45: all right, what would a reliable robot look like?

00:11:48: And where we ended up is this idea that you have a finger

00:11:50: and the finger has five motors in it and those five motors drive four joints.

00:11:55: That's a technique called N plus one actuation.

00:11:58: Sounds a bit weird. Why have the extra motor?

00:12:00: Turns out the extra motor means that you're never

00:12:03: going through a period of zero force on any joint.

00:12:06: So you never have backlash, which is really, really cute.

00:12:10: The movements become, the small, subtle movements become much easier to do and much more

00:12:15: flexible. So we have this modular finger design, we designed some new sensor control

00:12:20: architectures, some new buses, and we kind of built all that into this package. It's

00:12:25: bigger than our previous hardware. One finger weighs a little over a kilo,

00:12:29: okay, really got quite chunky options. But it's very modular. So you can swap a finger in

00:12:33: and out without needing to change anything in your experimental setup.

00:12:37: And how fast is the hand to open and to close?

00:12:42: It's embarrassingly fast. We have some video where we're just like, wait, what

00:12:45: happened? It's like, you can't quite see it moving. If you're running it in

00:12:48: tall control, it moves very, very fast. The joint control setting limits it to 180

00:12:53: degrees per second on any of the joints, which is which is an open and close in a

00:12:59: second without too many problems. Yeah.

00:13:01: And how many new do you handle or how do you do you provide for a fingertip?

00:13:08: So we were looking at this and saying to ourselves, do we want to build a reliable

00:13:12: robot or a strong robot? And okay, we could build a very strong, very reliable

00:13:17: robot, but how big would that get? And there's a certain limit which is going,

00:13:20: okay, this is starting to get silly. So the spec that we came up with was to be

00:13:24: able to handle, I'm going to call them typical objects, a couple of kilos in the

00:13:28: fingers. So three finger robot will hold one or two kilos without any problems.

00:13:32: And across the working range, anywhere in the working range, it should be able to

00:13:37: exert eight newtons at the fingertip.

00:13:39: Okay. But what is now your unique approach compared to other companies?

00:13:45: I think part of the, this is one of those projects where you go, well, actually,

00:13:51: there's a whole stack of things going on that reinforce each other in the design.

00:13:54: So we stuck to traditional actuation in terms of you just using top quality

00:13:59: electric motors because we knew that that would be a really good

00:14:02: approach. From Switzerland, it's always from Switzerland.

00:14:07: Absolutely. If you're buying motors and gearboxes and you're not buying from

00:14:10: Switzerland, please let me know.

00:14:11: Yes. But then in other places, we were able to go back and say, okay, what's the

00:14:18: right approach for doing this? And then often there wasn't a right approach.

00:14:22: So we said, what are the possible approaches? And then we build a test

00:14:25: program and work out what gave interesting information, what gave good data, what

00:14:29: gave good performance and go back to the researchers at Google DeepMind and say,

00:14:34: hey, we see that there are two or three ways we can do this. Which do you like?

00:14:39: And they would often say, we like all of them.

00:14:41: Okay. And then we'd have to go back and say, okay, how do we put all of them in

00:14:44: there? And that's why there are two completely different tactile sensing

00:14:48: technologies in the hand.

00:14:49: I can order different tactile sensors.

00:14:52: We have different sensors in different regions.

00:14:55: So if you look at your fingers, you'll see in the middle and proximal

00:14:59: phalanges, you've got this kind of fleshy pad that comes into contact with

00:15:02: things when you grasp. And there we cover those with lots and lots of

00:15:07: three-axis hall effect sensors with magnets. So really nice sensing technology

00:15:11: gives you a really good understanding of contact and deformation and force.

00:15:16: In a compact form, it can be made over the surface of the finger.

00:15:20: But then at the fingertip, we have more space available.

00:15:23: So there we have a new sensor we developed, which is based on a stereo

00:15:28: camera pair pointing at a pattern of markers on the inside of the finger,

00:15:32: which is squishy. And as you squish it, you can see these markers move around.

00:15:37: And we found that that's so sensitive that we've yet to have anybody

00:15:41: managed to touch the fingertip without being able to detect it in the contact.

00:15:45: But it's robust enough to go up to people pushing as hard as possible,

00:15:49: 70, 80, 90 kilos of load on the finger.

00:15:51: So newtons of load on the fingertip without it being unduly saturated.

00:15:57: They're very, very responsive.

00:15:58: How do you achieve this technical occurrence?

00:16:00: See of your hand.

00:16:02: So in that case, it was by looking at our options and trying them and trying them.

00:16:07: And sometimes we said, actually, we're going slightly too far here.

00:16:10: So we had a previous iteration of that sensor, which was so sensitive.

00:16:14: It was clear that it was going to be really hard to process the data from it.

00:16:18: And then we backed away a little bit.

00:16:20: So we just have a high frequency video stream running down the hand to to analyze.

00:16:25: But then in other areas, there was kind of very obviously a

00:16:27: after a couple of iterations, it was clear that there was a approach that was

00:16:32: excellence and we didn't need to go beyond that.

00:16:36: We could see clearly that was the right approach.

00:16:38: It gave us the data that we needed.

00:16:39: It gave us the reliability that we needed.

00:16:42: And we knew how to build it, most importantly.

00:16:45: You mentioned DeepMind as your partner, as your customer, maybe.

00:16:49: That's fine.

00:16:51: That's a research company, a huge one.

00:16:54: But Rich, how do you scale now your robotic hand?

00:16:59: Well, the nice thing about a deep research organization is that they will

00:17:05: often they buy a lot of hardware.

00:17:07: Yes. Yes.

00:17:07: So that's nice from a kind of purely purely technical point of view,

00:17:12: commercial points of view.

00:17:13: But what we're very interested in now is say, well, actually, what we've done

00:17:16: is we've gone away and quietly over five years, we built a new category of robot

00:17:20: hardware. What can be done with this new category of robot hardware elsewhere?

00:17:25: What can you do when you don't mind if your robot hits things?

00:17:28: What can you do when you've got the very, very high levels of dexterity and

00:17:31: agility that a human hand has with reliability built into it?

00:17:36: And what can you do with all this data that we're generating?

00:17:39: So we're now going out and we're talking to the community and saying,

00:17:42: how can we take this technology and how can we apply it into different places,

00:17:46: different areas, different classical robotics challenges?

00:17:50: Where so far no one's really used robots because they weren't quite good enough.

00:17:53: Does this technology unlock good enough for you?

00:17:56: So how do you see the development of the humanoid systems?

00:18:00: Is that the new trend you want to follow?

00:18:02: There are lots of people doing humanoid.

00:18:05: I personally, we started out doing bipeds 25, 30 years ago.

00:18:09: And we stopped because we realized that bipeds were too far away.

00:18:12: Now the world has changed.

00:18:14: And, you know, I think somebody pointed out to me that to build a

00:18:17: humanoid from scratch takes 60 engineers a year.

00:18:20: And it's almost a quantifiable at that level.

00:18:23: So we have some new technology.

00:18:26: Is it potentially interesting for humanoid?

00:18:27: I'm sure it is.

00:18:28: Is it interesting for anyone else in the robotics community?

00:18:31: Yes, I think so as well.

00:18:33: But you mentioned 150 sensors.

00:18:36: That sounds very, very expensive for, I would say, common robotics applications.

00:18:42: It's true that we have built, we've done what Shadow always likes to do,

00:18:46: which is to build a thing absolutely at the top of the mountain.

00:18:49: Yeah.

00:18:49: That's kind of where we always like to start.

00:18:51: Could, are there simpler things that could be built?

00:18:54: Oh, I'm sure there are.

00:18:55: Are there ways to adapt this technology to be more general purpose?

00:18:58: I'm sure there are.

00:19:00: What we found very interesting was when we ran the hardware through our pricing model,

00:19:04: it actually came out cheaper than our old hand.

00:19:06: OK.

00:19:07: OK.

00:19:08: It's just a surprise.

00:19:09: How do you do that?

00:19:10: It's just, it's to do with the fact that the engineering was designed

00:19:13: from the ground up for repair.

00:19:15: OK.

00:19:15: And that turns out that repair and manufacturing in the first place

00:19:19: are remarkably similar places.

00:19:21: So something that's very fast to repair is something that's easy to strip down

00:19:24: and put back together again.

00:19:26: And that means it's pretty easy to put down in the first, put together in the first place.

00:19:29: So do you talk to the, to our gripper providers like on-robot,

00:19:34: Schung or Schmaltz or is it, are they competitors and you don't want to talk to them?

00:19:39: Historically, we haven't because we've just been operating in somewhere

00:19:43: that's a very, very distinct location.

00:19:45: But now we are talking to more people who say, well, you know,

00:19:48: we have these products that we currently use.

00:19:50: How does your new system fit into that landscape?

00:19:53: And we're very interested to understand, you know, where we can contribute

00:19:57: and where some of the technologies that we've developed may be able to make these

00:20:00: like other types of hardware more versatile and more useful.

00:20:04: I want to come back to NVIDIA at the beginning because the CEO says

00:20:08: without AI and the design phase and they won't get any further.

00:20:14: What about you and your hand?

00:20:15: It's interesting that we didn't really use any AI in the design process.

00:20:22: But you collected during the testing, right?

00:20:24: Well, because what we were able to do was to essentially

00:20:27: take some very bright people and lock them in a room and say,

00:20:30: don't come out until you've got something that works.

00:20:32: That if you can afford that approach, it's a great one to take.

00:20:36: In terms of future designs and future systems,

00:20:39: I think one of the nice things is that we now know that we can use

00:20:42: some of the tools from the AI toolbox to challenge our designs

00:20:46: and to give us kind of things to explore for the next steps.

00:20:49: Yeah. But you are using simulation tools, right?

00:20:53: You will use AI based simulation tools.

00:20:55: We use the simulation tools, but they tend,

00:20:58: where's the weirdly they tend not to be AI based.

00:21:00: Something like Majoko or NVIDIA's Isaac tool is a great simulator.

00:21:05: They're very powerful and they basically work of classical physics

00:21:09: and just really good modeling of the classical physics,

00:21:11: which is why they're interesting for AI,

00:21:14: because they allow you to verify what the AI produces by testing it

00:21:18: in a real simulator, a real physics model.

00:21:20: What is now possible with your hand?

00:21:24: What are the plans of DeepMind?

00:21:26: What they want to do, they want to do reinforcement learning with the hand?

00:21:29: What else is possible with the hand?

00:21:32: I would love to know what they want to do.

00:21:35: Honestly, we've been through this really great relationship

00:21:38: where we built stuff for them and they went away and they came back

00:21:40: and they said, "That won't work because."

00:21:42: Okay.

00:21:42: They said, "Why not?"

00:21:44: And they said, "Because."

00:21:45: Okay.

00:21:45: And we're, "Okay, fine. That's fair enough."

00:21:47: They very much, they know what they want to do

00:21:49: and they have a strong focus on their particular innovations.

00:21:53: And we're quite happy to sit here and say,

00:21:56: "Okay, we're the hardware guys here.

00:21:57: We don't need to actually understand the rigors of your learning systems

00:22:01: algorithms, because what we need to understand is the rigors of your lab

00:22:04: where you beat the robot up and we make it survive, though, that damage."

00:22:10: I want to come back when you said we want to also tackle

00:22:13: different robotics application.

00:22:15: How do I integrate the hand into my robotics system?

00:22:19: Is it an API or is it a ROS-based operating system?

00:22:24: What is your idea?

00:22:25: It's at the moment, it's all in ROS-1.

00:22:27: Okay.

00:22:28: We have a nice open framework where you have joint position control,

00:22:32: joint torque control, tendon force control available.

00:22:35: You have all the sensor data published in a nice big ROS topic,

00:22:38: lots of ability using dynamic reconfigure to change the configuration of the system.

00:22:43: But it is an open API in ROS, in fact, it's an open API in ROS-1.

00:22:47: Okay, fine.

00:22:49: We've been checking that the entire project we've had moved to ROS-2

00:22:52: sort of in the middle of the to-do list.

00:22:55: And it's now the end of the project and it's still in the middle of the to-do list.

00:22:57: Okay.

00:22:58: It's not been that critical to do it.

00:23:02: You're a little bit late to the party, right?

00:23:04: Well, we have reliable stable running systems

00:23:06: that would run for thousands of hours on ROS-1.

00:23:09: We changed it, we have to re-qualify.

00:23:11: So that's part of the reason for sticking there.

00:23:14: Okay.

00:23:15: So what are your next plans with your company?

00:23:18: Well, we're very excited to take this hardware out and say to people,

00:23:22: "What can be done with this? Where can it go? What can we do?"

00:23:24: Where can buy it?

00:23:26: From us, from our website.

00:23:28: Drop us an email.

00:23:29: We're in production now.

00:23:31: That's a really nice thing about this.

00:23:32: We've actually transferred the production across to our Madrid team.

00:23:36: So you can, if you feel the urge, just drive down to Madrid and collect what you want.

00:23:40: Oh, that's a European...

00:23:41: Yes.

00:23:42: We hedge our bets there.

00:23:45: We hedge our bets there.

00:23:46: And you produce in Spain or where do you produce it?

00:23:48: Yeah, we produce it in Spain.

00:23:50: Obviously, subcontract manufacturing happens all over the planet,

00:23:52: but the final development assembly test is done in Spain at the Madrid office.

00:23:59: Rich, we wish you all the best with your new hand, with a new approach,

00:24:04: with your new key customer, DeepMind.

00:24:06: We learned that they order a lot of your robots.

00:24:10: More than most people do.

00:24:11: More than most people.

00:24:12: It was a pleasure. Thank you very much and all the best.

00:24:15: Thank you, Robert. Pleasure.

00:24:17: Robotik in der Industrie.

00:24:22: Der Podcast mit Helmut Schmidt und Robert Weber.

00:24:26: www.bergevlog.com