Podcasts›Soft Robotics Podcast›The case for three fingersLike💬 Be the first to comment →VideoAugust 18, 2026The case for three fingersFrom Soft Robotics Podcast, hosted by Marwa ElDiwinyVideo of Tacta Systems on fluidic tendons, conventional tendons, and direct drive.Play videoSaveWatch on YouTube TranscriptYouTube automatic captionsSearch the transcript0:00And so, for example, I'm showing this five-finger hand here. And so, you have a incredible amount of engineering that goes into how do you get the fluid all the way through these different sections and do it in such a way that if for whatever reason it fails, you can just pull it off and then insert the next one without losing any time on the production floor. And then the other part is that I want to make the fingers u very thin and and the reason for that is that u you know if you look at where we are going to go in the next you know 3 5 10 years we want to go beyond the human hand we want to be able to do more dextrous task more you know things that I wouldn't you know I don't have a good marketing term for it right now I wish I was a better marketeteer but we call it the tiny you know so the hand is smaller than the human hand we had to develop these these sensors, right?0:53And the we call them tactile sensors. And that again is a a completely new invention because everything you see today is either optical gel-based or you have you know what we call PO resistive you know FSR force sensitive resistive sensors. And I came back and I said when it became a there was a lot of focus it all came down to three fingers. The thumb is super important because I'm picking up a part. I'm moving it around. So that level of dexterity is very essential.1:24The part that really got me excited about this whole thing was well physical AI plus tactile becomes dexras. That's the missing link. The cinderel gap. So when you you hear discussions about tendon base versus direct drive is that tendons are very difficult to model. you know the cinderel gap is getting maybe a little bit better but it's still not like direct drive. So the big advantage of the direct drive is like cinderel gap is almost zero. how does your fluidic tendons compare to normal tendons when it comes to doing the modeling and then also on the sim to real gap?1:58How do you think about having glove that damage resistant or having maybe some capabilities to avoid being damaged? Well, Mara, we have a very special guest today, Vicram of Tact Systems. And like a little story to kind of tell is that we've actually met before in real life for the the very first time we met was in real life about two months ago in Detroit at Re-industrialize. And Vic wanted to show me a rather interesting hand design.2:29And so we hid behind someone's booth and he was able to show it to me. and I haven't been able to report about it until now because last week attack officially came out of stealth and it was quite interesting. So Vic, what what do you remember about that meeting? Yeah, first of all, thanks for having me. you know, we were trying to find a place to show the hand and it's a you know, it's a a gold colored hand so it strikes out when you look at it from you know, any different position in that conference.3:02And so I'm like, "Hey, let's go behind this booth." And they're like, "What are you doing behind the behind our curtains, you know?" but it was a really memorable experience. Yeah. Yeah. And, it it was something else to see because it's very different than everything else. So, so normally we we look at, you know, how you do the classification of all the different hands and usually it's like there's tendons and then there's like direct linkage and then you go like down the direct linkage tree and how do you break them out? It's like, well, you know, are you actually doing linkages? Are you doing direct drive of the joints? you know, all that and then you go down and everyone's like most tendons tend to be the same.3:35It's just like an argument of how many you need to have where they need to be, but tendons are tendons, right? and I think and then there's always like there's always like the outlier, some other way of being able to make a hand. And I think there's only like one there's probably a couple, but I mean the main one example I can think of is clone is using hydraulics. but you're doing something a little bit different. So it's fluidic tendons. and can you explain what the difference is between that and say hydraulics and normal tendons? Right. And so if you you know if you step step back I mean the cabled driven tendon driven hands u are really interesting u you know in contrast to the motordriven hands.4:15And so that's been really something that's been fascinating us for a really long time. And as we drilled deeper into it as you have you know outlined or talked about it in your podcast the tendons break and so we are kind of stuck with this situation where I can't use fine motor or you know motor u driven hands and then tendon driven hands especially in the industrial environments that we are trying to go after where you literally to be you know to be industrial grade you need to be dextrous you need to be fast you need to be very strong, precise, and most importantly, you need to be reliable.4:53Yeah. Yeah. Yeah. Yeah. Yeah. When you say industrial, as soon as you say industrial, right, you mean robust. Exactly. That's like number one. If it ain't that, I don't care about all the other stuff. It's got to be robust, right? And so, you know, given given myself and my co-founders, I mean, we come from the manufacturing industry, especially from the electronic side of the industry. And you know you are when you are making high value products every minute that you lose in the factory you know customers are calling you they're losing you know tons of money and so then the context becomes u the the conversation is not how many degrees of freedom do you have or you know how fast your robots can move and so on so forth.5:36It becomes what's your meanantime between failure? What's your meanantime to repair? What's your mean, you know, what's your availability? You know, how much time does it cost for me to service these robots or hands or or machines? And so the equation completely changes when you go into a you know, into a a value high value manufacturing environment. And so u the advantage we had was that we didn't have any you know religion on what type of robotic systems you want to use. And so you're always attracted by hydraulics.6:07U the cabled driven hands look very interesting. we know you can't use hydraulics in these kind of environments because if there's a leak for example the highly pressurized systems you have pump and valves. A lot of people have tried it. I'm not saying it can't be improved any further just like you know circa today. And so we came up with this really unusual design which is kind of a I wouldn't call it hydraulics. It's I think it leans more towards hydrostatics. And the idea here is that you have an actuator that pushes a piston and there's a fluid very very small quantity of fluid that then drives itself drives this a remote actuator so it's not on the hand itself and then it pushes the pistons that then you know move the fingers right and the thumb and so for example I'm showing this five finger hand here and so you have a incredible amount of engineering that goes into how do you get the fluid all the through these different sections and do it in such a way that if for whatever7:07Reason it fails, you can just pull it off and then insert the next one without losing any time on the production floor and not have all the issues of you know leakage and we call it vap rates or you know things of that sort. So u so today where we are you know these and the nice thing about the actuators are they become really simple right and you don't have all the complexity of I need to have this kind of drive or that kind of drive just by the the most basic actuators you can find and and drive it and you can either put in the chassis you can put in the shoulder you can put it in the torso the idea or even closer to the to the to the wrist but there's nothing that is there's currents in the in the hand.7:53There's no there's no heat, right? There's no none of that. So, so that's the general idea. This is to best of my knowledge, no one else has done this before. and that was the and it's a you know a significant focus because I mean a significant amount of belief that or conviction that this might work because everyone else is in order trying a different approach and there's a lot of you know expertise that's out there. So we had to kind of build a groundsup team, a mechatronics team that specialized in these kind of environments.8:30And we'd call people and say, "Hey, do you want to come and join us and work on this?" And they're like, "What are you trying to build?" And they're like, "I had never heard of anyone trying to use this before for a robotic hand." So that's the kind of team that has come together to kind of build this out. Okay. So if I guess one way of looking at it is is like using a syringe where you would like just push on a syringe the fluid would go out and then you can pull on it and then you get back and forth. So the one advantage of I guess normal tendons is that you can work them work in both directions not just in tension right but also kind of compression.9:03Now do you have like more power in one direction versus the other or is it the same? It's so this is actually dual actuation one for the push and one for the pull. Y so you have two lines that go in. So one is two lines push out and so u that this allows us to maintain power density and and speed at the same time. So for example each of the fingers can lift you know we have shown that we can lift up to 2 kg 2 kg of weight per fingertip and then and do it at about 360 or 400° per second.9:36So you can move this really fast. Okay that's efficient. Yeah. Yeah. Now the so what's the difference between normal hydraulics is it's it seems like you are actuating each one of the tubes with your own set of piston. I'm not sure what kind of actuator you're having, but obviously something that's makes the the piston go up and down really quickly. You you can get a lot of speed out of it if you have like a big cross-sectional area of that piston and small going out. You don't have to move very fast there and you get a lot coming out quick. versus your standard hydraulics is that you've got a hydraulic reservoir and you have a bunch of valves that you have to open and close really quickly to be able to get that control.10:12So, so that's what you've changed. Again, the nice thing is it goes there. Is it What kind of fluid are you using? Is it an oilbased fluid or is it Yeah, it's an oil based fluid. It's like a but it's very very low volume and because you can first of all there there's no leakage. It's a closed it's a closed system. it's completely sealed and that also is a signific significant amount of you know invention engineering that was required and and the and the and the and the chemistry and the physics changes dramatically at the at the interfaces.10:43You know this is not your hydro this is not a hydraulic system so you're not worried about the kind of things but on the other hand we are maintaining or we're measuring the temperature the pressures very accurately across the line and so when we see for example something is dropping you can shut the system down so there is no you know evaporation or leakage of the of the of the fluid and these are I mean you know we have spent a lot of time on optimizing the the the properties of the the the oil or you know call it the fluid.11:14Yeah. Yeah. And now one of the advantages of of of using this is like if anyone doesn't know hydraulic like just look at a backhoe. I mean they're just amazing because all the power of the backhoe is coming from the body. It's not anywhere on the arm. It's just through the tubes that go on out there and you can get remarkable power out of it and also remarkable control. I mean you can see what some of these operators are able to do. And so the advantage of that is is you're taking the power source and you're putting in the torso which means you don't have distal mass to worry about. So that's number one. Number two, one of the problems with tendons is that it has to go through the wrist and as the wrist moves you get a lot of cross talk because the tendons are going to stretch.11:53They're going to change length. They're just that depending on how you wire them. One solution which 1x has come up with is they're using like bowden tubes to try to do that which is basically like the covering on a bicycle cable. you're effectively I don't know whether they're technically considered bowden tubes but you are kind of doing that and what that means is the length of your tube never changes and you can care less about what's going on the wrist so that means but you don't have any stretching problem there that you would typically have even when you do run the tendon through at some point it comes out of the tube and then has to like make itself through the fingers.12:26Yeah. one one of the you know so so the the whole you know this is one of those you know we come from the manufacturing industry right I mean you know we have done a variety of different types of you know very advanced mechatronic systems the part that was of particular interest to me was and my team was that u you know how do I put joint encoders on each of you know on across we have three joint encoders right this might be there may be other people doing it but now I want to get a lot of precision And then the other part is that I want to make the fingers very thin and and the reason for that is that you know if you look at where we are going to go in the next you know 3 5 10 years we want to go beyond the human hand we want to be able to do more dextrous task more you know things that I wouldn't you know I don't have a good marketing term for it right now I wish I was a better marketeer but we call it the tiny you know so the hand is smaller than the human hand and so while it's congruent to13:26The training that we do to the the hand and you can find people with smaller hands to kind of you know to transfer do the training of those hands but you so because I don't have any heat on this on this section I'm can be very small I've got very small controllers I can reduce the amount of volume of the of the flu of the of the the fluids I can start doing things that we as humans today are not able to do and so this idea was very very important Because today if you look at you know the manufacturing processes for a lot of very high value products I mean today you know I and and it's it's an amazing time to be doing robotics for sure.14:07So u you know we have a variety of u robots u vision based robots u with you know more traditional hands or grippers that do a very efficient job in moving objects. Now you know each robot you can do your unit economics and ROI to figure out which suits your purpose but on the highv value side we have not been able to replace or actually complement augment or enhance the performance of the human labor and that's where this this this architecture becomes really you know exciting so initially start out with the existing work high value work and then go beyond and you know know for example if I'm trying to make a you know a u a device that goes on my ear or I'm making a pen or I'm trying to make you know heart valves or pumps and you know stances and all those kind of you know or things that I do in in motor what they call inspace motors you know things of that sort which15:08Require incredible amount of precision incredible amount of you need to be able to use tweezers and so on so forth and then of course you add the the requirement that they needs to last for a long time. So, going back to the reliability part, you know, an important part, remember MTBF and and I've done I've done so many ROI calculations over the course of my 30 years or 35 years here and literally customers don't care what is being used. No one you know no that you know someone goes and once you meet the basic performance you know of course there's a threshold that you need to be able to do certain things.15:44It just comes down to cost of ownership total cost of ownership. That's why I keep saying ROI is king. That's all anyone really cares about if if you can give them something that hits that. And of course, part of that whole calculation is is things like robustness and reliability. Now, from what I've seen is that some of the best tenant based hands out there might be saying that they're able to do a couple million cycles before failure. yours, I think, is it over 100 million that you're publishing that you've been able to do? Yes, I would say 100 millions. I would like to get there. but you know right now we have u you know you know we have demonstrated that we can work up to 30 million cycles per joint and the reason why we put that target was there 30 31 million seconds in a year.16:29And so you know so that that allows a kind of a a framework that depending on what frequency you use what your duty cycle is this is you know within the scope of what we think and what's the failure mode is it actually in the fluidic tendon is it actually joint is it a combination yeah so there was no failure mode there was literally I we crossed that threshold and we had a balloon in the company which said 30 million and Oh yeah yeah yeah exactly 31 million cycles takes you a long time.17:00Yeah, I think at that point it's hard to get data behind. Okay. Yeah, you got that point. Yeah, because because you know they say you know if you take a step back and what we unveiled was a tact right and the idea behind the tact is four parts right there is the hand which is very essential. I mean you know we humans are we have an amazing brain AI is catching up decent eyes computer vision catching up or in some cases beyond us the hands is the one part that you know you take the hands away all the high value human work goes away but then the second part of the equation was well I need to be able to touch and feel the real world and that was really important to us and that's hence the name of the company tactile right so and so you know The robotic hands need to be able to feel forces, normal force, sheer force, contact, defamation, temperature.17:56And so that became a the you know apart from the 30 million cycles. The other part we had to kind of invent was we had to develop these these sensors, right? And we call them tactile sensors. And that again is a a completely new in invention because everything you see today is either optical gel-based or you have you know what we call P poses you know FSR force sensitive resistive sensors and from you know my co-founder Nah who spent many years at Apple and has been is an expert in in in touch and force basically came up with this idea that why don't we use men'sbased sensors.18:40Now, think about making mems. So, at the company at Tacta, we make our own PZO electric MEMS. we design our own analog to digital converter chip. So, each sensor here, there are 256 sensors on each patch that go across the entire glove or at the robotic hand. This is the robotic finger that that I'm that I have here. And we have a 12-bit ADC that converts the at source on each pixel or taxel you call it, right? taxel from you know the the the analog to digital.19:13So, so that was also a significant amount of effort because you have to design the the sensors, you have to design the silicon chip, you have to design the entire architecture, then you have to make this whole thing stretchable because you know you're stretching at incredible amounts of strain, right? And and so and it has to go across the fingertip. It has to go across the you know the the different parts of the finger and the palm is different, right? That again is very it's a very you know the sensor architecture for that. the sensor is the same and so that was the second part that we announced that we now have enabled touch and force to this robotic hand and the third part which was you know equally important is how do you train these robotic hands because I don't have Shakespeare to help me out you know like I had in language and so we went and took the same sensor these are you know to kind of give you give you a perspective of how difficult this is each sensor is about 200 microns on side and it's about 30 to 50 microns20:13Thick and fully integrated with the with the silicon and each sensor is measuring 250 Pascal to 700,000 Pascal. All right. And the minimum pressure that humans can measure is a,000 Pascal. So we are you know operating at lower pressures. And then you have the you know one of the things that we have learned and I'll get to that is in the AI models are very very sensitive to the frequency at which you collect your tactile data because you are it's a full contact sport.20:49You're doing things at milliseconds. You're feeling textures spatial temporal responses of different surfaces. forces that come in sheer forces that are you know it's going to slip and I need to catch it like which I can't see visually right because I'm holding the object so we are operating up to about 400 hertz here and then we have imparted temperature for the first time in these sensors we are measuring temperature at about.1° centigrade resolution so human beings very good at knowing I'm hot or cold metal plastic I don't know exactly what the temperature is now I get a tactile ID as I'm touching this object So we took that sensor and the third part then becomes the the training glove.21:29I took the same we took the same sensor. So he and his team and they put that onto gloves. These are industrial gloves. These are bought at Home Depot and not exactly Home Depot but let's say you know as a euphemism or a metaphor right and then we have enabled the same sensor. So we have congruency now between the sensors on the glove and the sensors on the robotic hand. If it's 1 is to1 and now a factory worker you go to any automotive plant you go to any electronics plant right you go to any rocket or aerospace defense for cut proof you know woven gloves or bone and so now they can the you know they can collect data right and and now this becomes like a complete breakthrough because unlike tele operation and grippers naturally in the wild at very very low cost very high scalable feeling the same pressures.22:24Remember the the thin the these sensors are really thin so I can feel exactly what the human is feeling. and then you collect all this data. So let me give you an example. So we're working with a a I can't name the customer but let's say they you know they are somewhere in the US in the Midwest and they have a thousand workers doing fine assembly tasks. You get 2 million hours of data from that factory. Right now the same customer when they reshore from wherever they are right now in Asia they have tens of thousands of workers there that's 20 million hours of data so now you're getting very high quality scripted data u we have motion sensors on the back of the glove all right and those are very important because u we use IMUs you know because we tried we tried magnetic sensors and they're really good problem is you go into a factory touch bearing metal Right.23:16And so so you got congruency now on the tactile side from the human hand to the robotic hand for the first time. You got congruency on the finger motions from the the and then by the way there's a lot of work to be done on the IMU side because you know I call it the piano problem where you know go faster than gravity but then that's that you know depends on the application and then you have cameras right which are very important. So you have headmounted cameras. Yeah. So, so skill capture system, motion, tactile, vision, all included.23:49And and then you then you have the the fourth part that we talked about in our you know as we're coming out of stealth is that you have the dextrous AI model which is then got all not only the kinematics of the hand and the tactile sensor built into it. It then understands that you know the special sauce on the robotic foundation models is that you know you now have tactile in the VA or call it you know whatever the the the the terminology dour is world models or whatever they call it right.24:19So u we we kind of put that into that model and for the first time now I can get very high success rates with very u with significantly reduced training times right because the extra right so these are the four parts you know I mean it starts with a hand because you don't have the hand you don't have you know you can't execute right but then you have to u build all these four parts so the reason I went through this whole extended extended you know description was at 30 million hand 30 million cycles I'm like guys done now let's go and get the the the the pressure ranges right and then then the robotics learning team u you know is okay show me that you can actually you know when you're doing insertion tasks right if you're doing because or you're doing meshing tasks or you're doing those does this tactile thing really help I mean it all sounds great it's all theoretical ical until25:20You know you actually show that your success rates are improving. Okay. Now, just for the benefit of our audience, I would I want to summarize this this a little bit. First of all, you really took on the challenge of developing sort of like four technologies at the same time, maybe five, if I've lost count. So, it wasn't just oh, we're going to build a hand. You were thinking about what's the entire ecosystem that we need for that hand. So, the first is like how are we going to actuate it? So, like what you're really developing is fluidic. So right there you did it because that's the way you want to do it. But hands need to have good sensors and so you had to come up with the tactile sensors.25:53At the same time you think about how are we going to train this thing? We got to come up with a glove and we want to make sure it basically works hand in hand in a sense. Sorry for the pun, but to make sure that the the sensors you put on there are thin enough that when you put it on the human hand, it doesn't really feel like you've got something in the way. So the the human will still be able to feel it. It can also capture the data. And that's been a very difficult challenging problem for everyone else. and yeah, and so you got to get all those things together along with the AI.26:23And now of course they hand you you want to have the tactile. You want to still be a glove. But then the question is, well, how do we do the motion tracking? You could do it with an exoskeleton, but then it doesn't feel like a glove anymore. It feels like you're you're wearing a robot hand as opposed to a glove. And so using IMUs to be able to do that, which does solve some problems versus doing it with like RF or or magnetic because there's a lot of interference you can have if you're dealing with everyday let's say real industrial objects which are made out of metal. So if they're made out of plastic, those things work great, but then you start bringing metal in and suddenly you can't track.26:57But I'm assuming you are you also using like some of the egocentric and getting a skeleton on the finger. Now, now there's one Yeah. Now, there's one other thing that maybe we're breaking news here, but I seem to recall that when you you came out with a hand the other day, it was only showing three fingers. Am I right? Right. You just showed us a five finger hand there. So, what's going on? Yeah, I was holding it I was holding back u you know for you guys. No, just the the the reasoning behind this is very simple. I've over the course I mean you know my my co-founder Andreas u has been building things for about you know two decades more than me right and we have spent a lot of times in factories but after starting Tacta literally me and my team have spent you know days and months in Vietnam in in in China in in Mexico in all the factories that we can access in Taiwan Japan And you know and one of the things that came out was when we started analyzing27:57All the data that we had collected from our gloves right so I wear this glove and I have the workers right u actually do like a grommet plug or do a AI infra manufacturing or let's talk about laptop manufacturing what was completely revealing to me was that in every instance where there was a higher emphasis on quality quality yield which is basically first pass yield and tact time they were using three fingers.28:28Okay. So and right so using three fingers I'm seeing myself right here I'm holding a penoo three fingers cup of coffee well maybe I got a little bit there for holding it just for support but yeah if I'm taking this off a lot of times those other fingers are just going along for the ride. Exactly. And and what what is interesting is also remember the robotics fingers have one advantage over human hands. I don't get tired. So, I don't need to, you know, I have the grip and then because we don't have any heating issues or we don't have heat on our robotic fingers, I can grasp onto a a you know, 2 kg weight for for days, right?29:07I mean I don't have to so what what we what was interesting was I kept on looking at this data and and you know u and I have some really interesting stories that come out of Japan on why I need the the the the you know the ring finger which will be you know another time but the part that is really interesting is that after analyzed all our customers u you know data data on actual factory floors and that's where the difference comes between TLE operation and grippers right you have an exo skeleton.29:37I can't ask the guy. So, so you know, Japan has this really interesting word. They call it the takumi, right? The expert, right? And then the US we call it the the the engineer has the knack, right? So, I'm like, okay, let's get these guys and have them show me like over the course of a month of what or or a few days or a week in some instances. And I came back and I said when it became a there was a lot of focus, it all came down to three fingers. the thumb is super important because I'm picking up a part, I'm moving it around.30:08So that level of dexterity is very essential. And so I said, look, look, we'll have a we have a five-finger hand if customers want it. And it's modular if they want the five-finger hand. And there are instances where they might want to lift a heavy object and move it around and then have the same hand go in and pick up some small parts. Absolutely, we can support that. But remove the remove the complexity. go with three fingers and and the and and I keep on emphasizing the word I mean most of you guys can relate to that there's high value work and there's low value work and low value work is very important then I didn't say that neither is important it's just that high value is in the sense that if I make a mistake I will find out about it when the satellite is in space or the car is driving in extremely cold environments in Minnesota and you know the harness comes off Right.31:00So those are the kind of things I call you know very important. Right. So so simplify this to three fingers. but we have five fingers. If you want a six finger I can give you that too. If you want it's a very modular design is two thumbs is hard. that's the thumb is that I don't know if if you notice you kind of launch a spirited debate on X over like the three-finger hand and like you know suddenly you're seeing people coming out yeah you only need three like no you got to have five. And you know, usually the the reason for the five finger hand is the embodiment gap.31:31You know, everyone is just down to it's it's it's the problem. The question is how you training the data. are you going to be using like an uni glove which is always like three fingers, but it's it's kind of an odd three fingers because like three of them are kind of stuck together. you you get that kind of but it looks like you're not really part of the debate. It's like whatever everyone thinks is going to work you can do. And I seem to recall I thought for sure you showed me like the five finger hands. When I saw the three-finger, I was like, "Wait a minute, what's going on?" You know, sometimes you let the debate continue and I'm coming on your podcast and they'll see that and and the Yeah, maybe it's it's an unreal.32:07You don't have a dogma, which like you said, you approach the whole thing without like a dogma that's going to be tendons or direct. It's like how do we solve this particular problem? And when it came down to the finger count, it's like you weren't dogmatic about three fingers. You went out and collected data and it's like holy cow. It's like we only need three fingers from gloves, right? These are the gloves won't worn by the workers. Human data. Yeah. Build human data, right? For the first time, I can see that the the Takumi in Japan and the the guy the the guy with the knack in you know in Kentucky or in Texas and the guy is like literally touching and you can see like the the motion of the fingers and the and then he's like doing this really at very high frequencies.32:47He does something that RL picks up. You know, we use a lot of RL to kind of optimize our training model. And you can see these are happening at millisecond scales, right? Because we as humans have this tuned in that when I'm moving on this surface, I'm going to be feeling I'm expecting this bump, right? And because they're doing this again and again, when they get to that point within about a small window, which is why we went to 400 hertz, right? and and and so that temporal spatial temporal response is very critical.33:21One of one other important part that u about the hand which is really interesting is that this hand we actually make in Palo Alto. We have CNC machines right and you know I'm going to flesh it out a little bit more over time. It's not a it doesn't require any complex manufacturing, right? It's a it's something that can be scaled and scale is something really important. I mean that's one of the things that you know I I I always come back to is that we as a u we as a when we look at cost and and and scale they go hand in hand you know there's economies of scale you kind of bring it down but then the the other part is that it has to be the same thing with the sensors.34:04I mean we use the the the world's leading foundaries to make our sensors right so lot of scale and so this is a part of the part that also we struggled with early on like if I go with motors you know where's the supply chain u you know with tendons you talked about you know putting on wrappers around it like a bicycle wrapper and so on so forth so yeah so we are extremely customer-driven you know you know in a it's it's a perspective.34:35We are not coming out of a research lab or you know which is amazing right I mean the kind of things and we're all as an ecosystem the TAM is huge the total available market for robotics is huge one more thing I wanted to highlight which I've seen in your podcast being talked about I think more recently is the form factor you know what type of robotic form factor do we need in a in a factory right and one of the things that was you know is the first and I and this is I've never in my life hired a worker in a high value production facility for the legs, right?35:11You would look at how what they can do with their hands, right? And and their intuition and their and their neck. And so our perspective is that by focusing more on the hands and going for a modular design which could be on the torso, chassis, shoulders, you know, we are agnostic on the on the form factor doesn't have to be on a humanoid. We actually are launching with a cobalt, right? and and there are some specific examples that I can give you where I have tried robotic arms that are more humanlike and they will improve over time but the existing cobalt infrastructure is amazing it is tested it's proven it's so stable you have a variety of suppliers that have got incredible differentiation on kovartic arms so when you have a very stable arm and already already on a cost curve right and and in available in the US, available in other parts.36:05We have got some great suppliers out of Europe and we can go in and do part do you know every production engineering manager that I've talked to you know everyone has their audience there's a CTO there's a chief supply off supply chain officer our persona that we go after is a production engineering team they're the guys who deploy this and they're like Vic just get this on a coatic arm and get into the market right and then then we and then we'll tell you whether we want legs or we want a base and then we'll partner with people who make that.36:39Yeah. And so that's the approach that we have taken. So the form factor is something that is understated on our approach. we we feel like the the you know a production engineering guy has nightmares. He doesn't want to be called at 2 o'clock in the morning and something happened, right? And so the we take out u so everything on our on our tact bot is you know we last week was the technology unveil the product unveil will come you know in a few months and in that particular scenario you'll see you know giving a preview is that it's just a coart with tactile hands on them everything is off the shelf and that screams robustness right there because you can come up with a really robust hand but if it's on an arm which hasn't been proven then you don't know and you don't want to be that is the failure po point or the failure mode you have there.37:33And I've I've pointed this out many times on on different podcasts that the business end of the robot and of the person is the hand and the arms and the legs are all about getting the hands to where they need to be. And so you you're right, you're not hiring them from their legs. You're hiring them for their mobility and their dexterity. And what that mobility is mode is not really that concerning. now another thing I just want to kind of summarize for everyone which you you talked about earlier when you get down to talking about tactile is everyone is familiar with a pixel and they may be familiar with a voxil where the idea is a volume is like a pixel is like basically three-dimensional.38:11It's a volutric pixel or space or cell. but then there's tel and we should start getting used to that term because it's a new term that's just starting to float out there. A lot of people haven't heard it before, which basically means the tactile cell. Another thing is that everyone's measuring and unfortunately they're using like SI units for the measurement. And most people do not like SI units when it comes to sensing pressure. A lot of them will actually go empir go back to empirical units because PSI makes more sense.38:41And the reason is when you hear a Pascal, that's basically a Newton spread over a square meter. Okay? And a newton is like one kilogram. So imagine the amount of pressure that 1 kilogram does over a square meter. That not a whole lot, but it sounds like a big number when you said 250. It's spread over a big area. So again that you're talking about something which is about a quarter of what humans can actually sense. And we also even though our reflexes are only what about a tenth of a second something like that when it comes to measuring a touch we react much faster.39:20I believe it's it's more than 100 hertz. Is it close to milliseconds? Yeah, I mean it's there there's some really good u you know work that's been done over decades, right? I mean there's a group from Sweden I believe that actually has some of the more fundamental work that we had been looking at more recently. Was that out by any chance? I think so. Okay. It might be that Gustav's familiar with it because he's from up in that neck of the woods.39:50Yeah. And then and then Rodney, you know, I mean we started the company three years ago. I mean Rodney is a huge I'm a big fan of Rodney Brooks and you know a pioneer in robotics and I mean literally last year in September he literally came out and said guys you know you need to look at touch and we're like thank you Rodney finally. But you know the language the language I I'll go a little bit further. So if you take where we where we were many many decades ago or even a decade ago in terms of pixels or voxels right this is going to be a completely new language this is we are at the you know tip of the iceberg on or let's say starting point or the you know of the of where tactile is you know when we're naming the company we are like well I mean the mechanics of the hand are fundamental right that's the end what what is something that will where do we see this being really really futuristic and the AI models which are grounded in you know which will be need to be grounded in physics for physical AI.40:50So we call it dextrous AI and then everyone asked me what is the difference between dextrous AI and physical AI. So physical AI to me is spatial awareness. I I I know where I am. I need to get to a location. I need to know what's around me. I need to be able to move things. So in the verbs that we use in describing a what call a standard operating procedure or process in a factory floor, you have like verbs like I got to pick, I got to move, I've got a place, I've got insert, mesh, thread, so on and so forth. The part that really got me excited about this whole thing was well physical AI plus tactile becomes dextrous.41:26That's the missing link, right? and and so the dexterity part is heavily indexed on what I feel on these mechano receptors. There are four different types of mechano receptors each one serving a different purpose on the human hand. And essentially the only way I could reproduce that you know nahed my co-founder and Andreas my co-founder the two of them kind of got together and did this is and in temperature right is on a on a MEMS's device because the traditional FSR or PSO resistive sensors I can't do that with optical sensors.42:01The one of the things about optical sensors I want to kind of highlight a big fan of those. They're really really good. The problem is I can do that on this I can't do that on this right and so anyway so optical sensors will also give you very high frequencies you can collect data on these defa deformable but then you have to kind of bring it back within that eos within the system so we have seen in our in our u models that we see a difference when you start in specific use cases which happen at very high frequencies you need to be operating at about 320 20 Hz otherwise you'll miss that signal.42:38Okay. Okay. Yeah. And now if you're running at that speed, does that mean you're thinking of having like an S0 layer right at the hand itself? So you can process at that that speed. Yeah. So the the the entire the so so that's again the the the it's going to be really exciting on how we move forward over the next you know three I mean currently we have the execution of the bot is constrained by some of the other parameters the controller the the the kinematics of the hand one of the things that happens when you do dextrous manipulation if you look at some of the videos that we have out there and we'll put out more a lot more is that you're not moving your arms you're just doing a lot of things like this so that now opens a completely completely new control system paradigm that if I was to be able to do something at the edge and then use that to be able to and that's coming I mean that that does not exist right now but that's a real edge compute yeah you you'll see like for example I mean I am putting something in and I'm just moving you know my there43:39There are instances where I found that the the experts human experts are not even moving the wrists they're just doing everything at at a very very precise scale right and so those the kind of things that we will we will get to and I'm hoping that when we launch the tiny hand down the road that will be a part that edge the edge compute will be a very key part of that. Yeah. And and you can you can see other things that that while you may have like a high level command take the bottle and move it over to here. So the all the robot cares about is like moving from here to there and it can process it at a very slow hertz.44:14But when it comes to actually holding on to this, you want to make sure it doesn't slip out in your hand. So, the robot arm isn't doing the reaction. The fingers are doing that. And that that can be doing it at a very high update rate just to make sure it doesn't slip out. Which gets me down to your tactical sensors. Like the biggest thing is sheer force. So, I assume you have the shear force in there. The the other thing I find very important is that people may wonder, well, what's what's the importance of temperature? Is it just to know what something is warm or cold? And what people may not realize is that our fingers don't so much detect temperature as they do heat transfer like a rate of heat transfer and then from that gets kind of inferred as like if there's a lot of transfer going on there must be a lot of temperature and some I think is important right on the same finger I can feel this is a plastic and this is a metal.45:02Exactly. Because cuz that's that's when you go to plastic, you're not really feeling energy being absorbed or taken away from your hand. As soon as you feel something metallic, you know that. So that means that's telling you something about the thing that you're picking up. You don't really care about the temperature. You just care about the fact that it's more of it's got more of a thermal property. It's a thermal sync in a sense that gives you that additional information, which I think is again everything comes down to more modes. The more modes of information you have, I think you get quicker and better convergence on your training models.45:37When you start excluding stuff, I think it's much harder because it's it's lacking some kind of context or awareness of the situation to know the proper way of responding. So, it's got to in a sense try try to prepare itself for any contingency because I really don't know enough about what's going on. But the more I know what's going on, the better. And the other thing I think you notice is that you're kind of right is that whenever we focus on something, we tune out distraction and we really wanted to be focused in one area.46:07And that might be why you're right is that people are using three fingers because if you try to put focus on the other fingers, you really can't get what you need there. Then I think that's a very interesting observation that again updates my priors a little bit about the thoughts of what are going to be the ideal embodiment for hands because this is not dogma anymore. that this is actually going out collecting empirical data and then being able to come up with some kind of conclusion pure evidence. I mean like literally my robotics learning team came back and said I'm like guys you are the end you you know if the system fails you know and they're like no no Vic I'm telling you that this is you know we are seeing very very clear evidence that okay so the one one one other part about the tactile part is that we like I was going back to my like computer vision there is going to be a paradigm in the next you know 5 10 I mean I mean my I started out as a material scientist very early on in my career and so I'm always very fascinated47:07By what's possible with materials. Today I'm you know we're talking about systems and learning all about different things but you know I look at the the the the science that's going to be possible once people get excited about this is what is you know this is what we want to do even though we have been able to create receptors combined into one one sensor one you know the 200 microns by 200 microns imagine 3 5 10 years from now that tactile capability is going to be so much more advanced because we are going to have people who are going to start introducing other modes and then going back to my basic thesis that I already see with the the future of my hand right when I say I had the tactile hand this hand is going to be smaller it's going to be very different same thing with the tactile sensors there's going to be all these different modalities that we're going to start introducing that will make certain tasks better it'll make certain tasks possible and so you know that's the other and of course48:07The AI models Right. That's the other part that comes in. the the the tactile sensor is so interesting to me. I was stepping back and I was trying to think about something that came to my mind. For the first time in the next 3 five years, we'll have the complete repository or of all the human skills in a catalog, right? A full repository. we'll have people wearing these gloves they're wearing in factory today. there'll be new types of gloves.48:38I mean, you know, there'll be new types of embodiment and they will allow us to be able to capture why Vic is a horrible guitar player and you are a great guitar player, right? And then and then it'll you know nurses and then how is it that you know people are such amaz you know give you know are able to do certain things that you know and so we'll have a repository and you know we'll have the vision we'll have the motion we'll have the tactile we'll have other modes that I've not even you know I can I can imagine right now as we start enabling tactile so tactile is a whole new language right it's a like you know it they will people will talk about tactile like they talk about computer vision Right?49:19And you know it's going to be this very very complex part that comes in and that repository I talked about 30 years from now 50 years from now people will come back or 100 years from now we'll say oh that's how that particular you know that's the right way to play it or this is the right way to make it and then kind of build on that. And the thing is that experts are usually the worst teachers because they don't really know what it is. is it takes someone from the outside to be able to figure it out and just imagine the learnings from that is that eventually the AI might figure out this is actually the way to teach people and my father was an industrial engineer and he had the same problem is that they could always identify the person who had the knack and then they tried to get the person who had the knack to try to explain other people and they they just couldn't.50:05So he had to observe them to figure out what that knack was so he could then start teaching everyone else exactly what that is. And so yeah, don't worry, I'm a worse guitar player than you and I haven't even seen you play guitar. Okay, so I can tell you that for sure. and and again, I think I need to correct something. It's like I get my Pascal thing wrong because a Pascal is a Newton per meter, but a newton is basically a kilogram meters per second, which means conceptually we think of one kilogram is giving us about 20 newtons, right?50:37Or sorry no that that that would be like 10 newtons because of gravity. Okay. So in other words, yeah it's it's even lighter. I mean 250 newtons or 250 is like nothing. The description of yeah the description of of pressure. The reason why we went Pascal is that u it's it's a more universal you know language and then it takes out this whole complexity on what area are you measuring right and so from a from a from a from a you know the the shear the the normal all of those components are actually you know more important but you know this like what what is happening at that level and especially when you know where it is so it it's one thing to know the shear, but to get an idea of where it are and how that's the other part, right?51:29The Yeah. So, the first time, so now now that I've got we've got 256 of these per per you know, per patch and, you know, over time it'll go across, you know, it's going to go across the entire surface of the glove. I know the position where the sensor was, you know, where exactly the force was applied and how it transferred over, right? Because now we are I mean the the part that and this is the the you know talking about the future of tactile today we claim that there's between.8 and 1.2 mm is a spacing of the most critical of the fast adaptive sensors and then you have the your FA and SA I think they call it fast adaptive and slow adaptive sensors or mechano receptors.52:10We get to this point where today we're at 1 millm right and so the architecture that we have come up with allows you to go even finer right and so and then add more sensing capability once you go into MEMS now you start creating all kinds of new structures that I can start you know enabling which I can't do today right which I could not do with the other sensors so our our you know thesis is that depending on what the application is and which markets we're going into, we have now they created the ability for our customers to come and ask and and we for to measure and say oh I need to be you know I need to increase my fidelity in this area and now you can go on that you know on that chart on I this is that the architecture I can create that's the direction we're going I I think one other question This might53:10Be my last one. is like this the cinderel gap. So when you you hear discussions about tendon base versus direct drive is that tendons are very difficult to model. you know the cinderal gap is getting maybe a little bit better but it's still not like direct drive. So the big advantage of the direct drive is like cinder gap is almost zero. how does your fluidic tendons compare to normal tendons when it comes to doing the modeling and then also on the similre gap? Yeah. So on the on the model side, what we have found is that because our actuation is literally a motor-driven actuation, right?53:46So the the physics are driven by motors, we don't have any elongations, we don't have any distortions, we don't have any u u you know plastic defamation of the tendons and and so on so forth. We're able to model it based on the the actuators that we use, right? Because you're getting a lot of precision from there. the part of the part of the the the more interesting problem that we are trying to solve is that teley operation has a advantage in the you know sim to real right because you see what you see the grippers the exoskeletons u while not applicable in real world environments allow you to be able to mimic you know the the kind of you're basically mimicking the robot right you're kind of getting that and so what we are seeing is that insertion of some tele operation data insertion of the more of the mechanics of the the arm and the other parts they don't need to be the primary source of data54:46The the the corpus or the bigger data set is going to come from the human set but to be able to reduce that sim to gap we have found that you know those kind of things help so as we deploy further into the you know as we deploy more bots we will find that for specific speific tasks that gap is a lot much more addressable than because there are other parts that you know I've heard people talk about well you know the human is wearing the glove and the glove is not exactly you know it's the you've got a woven fabric and it's going to move but that's the reality is that the human hand if you look at how it operates over a certain duration of time which by the way goes back to your point where I don't I know I want to know what the expert worker did but I want to also know what the new trainee did right and and that gap is what is getting captured into the into the system so that when the policies come out it's like much more55:47Constructive seem to seem to real I mean there's a lot of work that's going to happen over the next you know 1 3 10 years I mean like this is going to get better and the exciting part is that you know our approach is a partnership approach we are agnostic on a lot of things on our ecosystem and Our expectation is that you know we we are for example compatible and and with any robotic foundation model company. So right we are compatible with any body that can absorb the the the hand.56:18So that I think will allow newer types of application specific ecosystems to kind of emerge and and it's such a huge market. I mean that's the part that I think you know I think is really exciting about this space. Yeah. I'm I'm done. Yeah. Go ahead, Marva. I think the one thing interesting and we saw in the I think Scotty covered one of the book the glove by one of the demo. I think it was 1x they have cuts in the in the glove and I'm not sure how do you think about the material embedded the glove and you have the sensing already.56:49So I'm not sure how you think about having glove that damage resistant or having maybe some capabilities to avoid being damaged. I'm not sure how cheap the manufacturing cost how the sensing. So that's the first part. and the second part because again began discussion about that you have a unique approach for the actuation. Do you think about artificial muscle direction because this is a yeah I'm not sure you're also planning to have a benchmark against like tendon direct direct drive when your approach right yeah I mean I I look at so let me answer the first part of the of the gloves.57:26Our entire approach is that so literally I've gone to our customer sites and taken all their gloves right and and and and seen which ones are the ones that are meaningful. Now, for example, there are certain types of like I call it like oven mitt type of gloves that also used in the in the factories, right? They're picking up like things that can really cut through their through their hand. whereas through the glove, most of the dextrous work is done with with u with thinner gloves, right?57:58And this is the like exa exact version that you'll find u the different form factors and so they're they're always using gloves in just about everything or there situations no you'll find yeah I'll give you an example there's one scenario where I've seen workers in Asia cut their the cut the the the tip of the finger of the glove. They want to get even more dextrous feel on antennas and you know the cables that they want to feel. So what was exciting about doing the test with those guys was that I said okay you're wearing our gloves you know you c can you feel that that that that pressure and even though I can't replicate exactly what she does my sensors are replicating what she can do right and so the the the robotic sensors can do that so in terms of the robotic sensors with a tactile because they're congruent right that that the congruency is very important by the way when you do RL everyone finds out that sensor one, sensor two,58:58Same you know the that matches. the cost part is very f is very interesting. The most significant part of the cost is the actual glove, the woven fabric. The you the the approach you're taking is that the the we get millions and millions of sensors from a silicon wafer, right? Or for whatever so the you're you're talking about I've got at in the most extreme case I might have 2500 sensors, 10,000 sensors.59:30You know, we're starting out with a smaller quantity. You got 1 millm spacing on the fingertip. you got 5 mm on the on the on the palm. When you go to the rest of the body, it becomes 20 mm. So the spacing becomes so the number of density of sensors goes down. So at the end if you look at asmtote and volume which is what we are designing for your most of your cost is going to be the fabric and there's a puck that's on the back which you know is a Wi-Fi based puck or could be a private network cellular network which is a direction people want to go in the future and so you can literally go into someone's home and start collecting data right and and and the gloves will wear out I mean there's no doubt that you know there's wear and tear The part that is different from the FSR or the PO resistive gloves that we see that people are showing today those wires will wear out over time right the specifically underneath the the PO the force sensitive material1:00:30Will wear out ours is an inorganic material it's basically you know silicon based or you know silicon type of material the sensor won't wear out the wiring inside will wear out and then that part becomes the the equation that We are constantly evolving. the gloves need to come in different sizes. That increases some complexity. we have I mean I I forgot to mention my other co-founder David. all he's done you know in the last 25 years is done integration of I wouldn't say all most of what he has done is integration of electronics into textiles.1:01:06So his team I mean if you come to TATA we have we have CNC machines and we have the same amount of space for textile integration right because we are you know and textile integration is very complex because now you got you know you got wiring on the back that needs to constantly move and so you got all this the wiring interface you got the the wiring on the on the glove itself. So you have to be moisture resistant. so all of these things u you know become really important as we go.1:01:37The good news is that thanks to all the work that's being done in consumer apparel and their integration of electronics amazing space. I mean if you are someone who's interested in that category of electronics lots of opportunities coming down the road and and that brings me to one other point first about gloves. My gardening gloves wear up pretty quick. So, I mean, that's that's the reality of gloves is that they don't last forever. and it's a consumable and we can make them cheap enough.1:02:08So, I don't think it's a big deal that they'd be wearing out so long as it's inexpensive. The other is compliance. So, it seems like by using the fluidics, you've got a certain amount of built-in compliance already and I assume kind of back drivability if you need that as well. And that is that's something that it's nice to actually be part of the mechanism sort of an analog compliance as opposed to something that you have to do in software, right? Yeah. I mean, you know, it's a I forgot to answer your other question about the other types of you know, muscles, artificial muscles I think you mentioned.1:02:44Look, I mean, we're constantly looking at, you know, the original hand we had actually had tendons on them. Mhm. They were combination of tendons and we did the same thing that we put, you know, the the the casing and we tried to optimize it and you had then we got down to one one cable tendon and everything else was fluidic and that was was failing after 150,000 cycles. Oh, and you can't even make a million with that. Okay. So, you're already at 30. Yeah, that's the thing that's really tough with cable driven tendants is you're lucky to hit a million, right?1:03:19So the meat meatronics guys are like vic we need to get this off and so and we were so I think the the the there's a the part that I think one of the things that most hopefully your audience and you guys will take away from this is that we have no religion. We are literally trying to solve problems. and that I think is because the the team at Tata is all about you know our customers need a solution. They want to be able to take a a legacy laptop manufacturing line that has not changed in 30 years.1:03:57They want to be able to make small consumer products that require a minimum of 900 human beings to work on it. and and then you have the entire challenge of automotive where stamping PL welding painting has already been automated but all the final assembly is still done by humans. So that's the part that we are u you know focused on customer-driven application focused high value work and and you know that's the audience that we are going after. So so I think the way to sum it up is that there's an expression out there fall in love with the problem not the solution.1:04:35And and that's what is that you you're you're embracing the problem. You're trying to find solutions and if a better one comes along, you're not going to be dogmatic about it and say we're sticking to the old one. yeah, that's a great chat. Mara, do you have any any final thoughts? No, it was really great. I think I appreciate you guys have different approach. I think this is unique, you know, Scott. Yeah. Yeah. Yeah. And and I think again Maro, you can break the fact that there actually is a five-finger version of the tactic hand. the you know the human hand is an incredible machine right and it's like you know that's the part that we all and if you go back and look at u you know the some of the more impressive machine companies in the world I mean ASML is a great example you know and it you can't think about the form factor you have to think of the the the what what's needed like how do I solve this problem and and we you know are super excited about this idea a that1:05:35There are other people have emerged. I mean your podcast is amazing. I mean I hear like all the amazing companies that come out and each one of them is kind of building into that ecosystem and solving that. So thank you for the opportunity. Well, thank you. I mean this has been been quite informative. It's always nice to to learn how other people are trying to solve the same problems. And and and now next time I you know when I I see you at reindust which by the way is a really important part of our mission to reshore and reindustrialize manufacturing in the US and Europe and other parts of the world.1:06:09We will not have to be behind a booth. Yeah. Yeah. I'm hoping that next year next year I reize I'll have a a bot you know said you know for sale. So great. Great. Yeah. So, I'll bring my checkbook. Yeah. Thank you so much. Okay. So, take care. Yeah.
