Tesla Optimus & 1X NEO: The Prototype Problem

[00:00:00]The hand is where the mind meets the world. And suddenly we can't just stop talking about humanoid hands in the expansive world of robotics. The 1x Neo hand was unveiled as the most advanced hand ever. But we still have a few lingering questions. Our very own Maharat Farimmani has slaved over the weekend to produce a definitive report on the state of data collection for robotic manipulation. We've got a spicy analysis of the real problem with the Tesla Optimus unveiling from our very own bots inside Scott Walter. Also joining me on the show is Dr. Gustav Anderson who I trust will keep us all in line and a special guest today Andrea Esposito from Foundation Robotics previously with the Tesla Optimist and Tesla Energy Teams. Welcome gentlemen.

[00:00:47]Great to have you all with me especially Scott on the train to Berlin.

[00:00:51]>> Nice to be here. Welcome aboard.

[00:00:53]>> And a special welcome to Andrea.

[00:00:55]>> Good to see you. time that he's joining us. Hope to have you on for a lot more.

[00:00:58]>> Hope so too.

[00:00:59]>> All right, so let's begin. The first item on the agenda today, a very interesting thread from you. Three questions I'd ask Burnt as an engineer who worked on the Optimus Hand team and current and lead at foundation. Throw a little shade there, won't you, Andrea?

[00:01:15]So, first degrees of actuation. 1x tech claims 25 do you count 20 go on to say PIP and DIP flexion appear coupled in the videos which is why the hand defaults to a pinch grip rather than a flat grip. A flat grip is inherently more stable especially on small objects.

[00:01:34]It maximizes contact area and uses the full pulp of the fingertip. I'm sure Dr. Gustav Anderson will have a bit to talk about the pulp there. And you say the thumb architecture is an interesting choice. 1X included thumb pronations, supernation, aial rotation, but skipped abduction, adduction, which is the most common choice in humanoid hands. I loved how the Wuji global hand solved this.

[00:01:58]You say it keeps abduction and abduction, then gets some axial rotation for free by angling the MCP IP flexion axis relative to the CMC flexion axis.

[00:02:09]Take us through your thoughts. Someone roasted me in the comments because I didn't ask any questions actually and that was funny. But um yeah, so I I analyzed the video of the motion of the hand and that's where I got my degrees of actuation count from. Um every time I saw PIP and DIP they were moving together. This uh could be because of um uh inherent coupling in design or poor controllability. Someone who claims that had that they have uh worked on the hand um uh show told me that that the uh degrees of freedom are actually independent. If that is the case, it's a matter of poor controllability. And if you guys want, we can dig deeper into that in a second. But in terms of motions again what I could see from the video was PIP and DAIP were moving in a coordinated fashion. So this joint and this joint move together. What that uh means and what that leads to is the inability to perform this kind of grip.

[00:03:24]So we tend to default to this kind of grip instead and we lose the uh ability to manipulate small objects by creating an extensive uh contact between the fingertips and the object. Um they do have a nice um palm flexion degree of freedom uh in a way that is better than many of the uh other hands in the industry. So they can effectively um create a sort of like ball space to wrap their hands about by uh flexing this degree of freedom. And then like you said, Roden, uh the thumb architecture is really interesting because most hands um have this degree of freedom which we call CMC yaw or CMC abduction and abduction and this degree of freedom which we call CMC flexion and extension.

[00:04:26]Notice that this is different from this.

[00:04:29]It's uh movement about this joint down here. In humans, because of the shape of the CMC joint, as we abduct, there is a little bit of an axial rotation of the thumb. And Gustaf can obviously comment on that uh much better than I can later on. Um but so most most humanoid hands neglect this partial rotation and only focus on ab and flexion extension. Um what 1X does is interesting because they have flexion extension and then they have axial rotation about this joint.

[00:05:09]Again, someone claims that uh there is AB ad there, but in the videos where they show the motion of the hand gloveless, I couldn't see it. Um I think I this still like the the Vuji approach still remains my favorite because the architecture remains pretty simple. Uh so they don't have this axial rotation but they get it to an extent for free by offsetting the angle um of flexion sorry the axis of flexion at CMC and the axis of flexion at MCP here and what they get is a palm that fold is a thumb that folds uh more towards the thumb than it would otherwise be. And I think in the demos, specifically the demo that they have made with uh Genesis AI, they demonstrate that that grip is really functional.

[00:06:06]>> Yeah, I kind of concur that the the idea of taking the angles of the two joints on your CMC and not necessarily making them perpendicular and nice and flush, which is like what the way engineers would want to do it. You end up getting that nice kind of motion, which is closer to what you would typically see for a thumb. Now, I I think another thing we need to back up a little bit is that when we first saw the teaser last week, I was like where we couldn't see the actual hand movement, but we were just looking at the actuator stack. We could clearly see what looked like indications of 10 or maybe 11 actuators that were facing us. And from it, we could also count a lot of the tendons that were coming out and it was somewhere around 20 or 22. You could kind of clearly see that was there. Um, but I surmised that just using symmetry, there could be more actuators to the back, which meant there could have been maybe up to 20 actuators, maybe up to to 40 tendons. And Baron dropped a comment in there, something about more. Okay.

[00:07:04]Um, and so from some other independent reporting and and others, the actual tendon count might be 44. Now, this is excluding the wrist itself. So, let's take the wrist out of the equation and not talk about 25. off but talked about the 22 and the other digits. It looks like they might actually be 44 and 22 actuators and 22 actuators could fit. We just haven't seen them. You know, we've only seen it from one view. It'd be nice to see a view around the back because I'm having a hard time actually counting 44 tendons. Again, it could be that a lot of things are blocked in the back, but you know, if that is the case, then it really is 22 degrees of freedom um because you've got 22 actuators there.

[00:07:45]And the question is, what are the joints that are there that come up with that 22? And since we know there's already the fifth metacarpal in there, that usually jumps you to 21 right there.

[00:07:56]Usually, everyone has only a two do CMC, which means your hand is 20 do without including the the pinky or the fifth metacarpal, which gets you 21. So, they have 22. That means they might actually have a five do.

[00:08:10]Do you think they count the wrist movement into their 25 or did I misunderstand you there Scott that they >> I don't know I don't know if if if Scott was counting the wrist but the tendon actrated wrist is in my opinion a very interesting um architecture I really liked it um I did some other deep dive uh posts and created a small prototype in uh in our office to just understand better how it was working. It's essentially a double differential capsson uh drive where the they have two very pancakey so very flat actuators sitting back to back and each of these actuators uh both acts on what we call uh wrist flexion and extension here and radial or ulner deviation in robotic term. um wrist fl wrist pitch and wrist yaw. Um and it's the it's whether the two actuators are moving concurrently or in opposite direction that determines which one of the two motions uh prevails. And so it's a way to uh make a very compact design very efficient from a transmission standpoint. It has some builtin compliance which helps with preventing damage of the actuators in case the hand hits a table, in case it hits an obstacle or or anything like that.

[00:09:52]Um, it allows to quickly like not quickly but easily adjust the transmission ratio so you can trade off speed and torque at the at the wrist.

[00:10:06]And so I overall liked that architecture a lot for the wrist. The only downside which they also have on a lot of the finger tendons is that everything acts in closed loop. uh meaning that every actuator is birectional. It can if it rotates in one way it pulls one tendon.

[00:10:29]If it rotates the other way it pulls another tendon and then the two tendons um travel to the joint and connect at the joint. This closed loop architecture is such that as over time you have either slippage uh meaning the tendons slip or creep of the tendons meaning they elongate you lose controllability because the um the path that the tendons now travel is longer than the path that it gets to get to the joint. And so there's there's something called backlash. Meaning that even if the motors are commanding a fixed position, you have a bit of uh unwanted movement um at the joint. That's the main downside of this architecture.

[00:11:18]>> Yeah. I just wanted to ask you Andrea uh based on your personal and professional opinion, how much faith do you have in tendon drive for h for Dexter's hands in the long term?

[00:11:30]So um I think I used to think that you need a tendon tendons to get to a high degree of freedom. The Vuji hand which I have seen in person in China last month disproves that. Uh so they uh have a very high dexterity with an architecture that purely um puts actuators inside the joints. Uh what that that architectural steel doesn't buy you is strength.

[00:12:00]Strength is the last big differentiator of um tendon driven hands because physics only allow a certain torque density. And if you want uh a strength that is comparable or superior to what humans have the volume the sheer volume that you have in the hand is not sufficient. you have to move somewhere where there is more uh more volume which in this case is the forearm and it's also the biological adaptation. Um so if I may ask a follow up on that uh then how do you compare it with a hydraulic hand for example because there you can have a high torque density basically.

[00:12:40]>> Yes that's valid. I think that that architecture is even more complex than the uh tendonbased one. Um because you because of all the pumps, because of needing to uh keep the system in in in pressure, because of the disaster that can happen in case you have a leak and and so on. I think like the the big unsolved uh matter in tendon-based hands is survivability of the tendons. uh you know that's what all the companies uh that are working on tendonbased hands including foundation are working on and developing solutions for um I do think that at a mature state of this industry where we want very high performances from robots including strength including um you know good thermal performance where the hand doesn't overheat which is one of the problem that Vuji has Um the tenantbased architecture is the uh endgame like the end the tenant based architecture can win the end game. Uh but there are a lot of reap of benefits to be reaped on the way there by uh more uh by simpler architecture like like the one that VHI has.

[00:14:02]>> Thank you. I mean I wanted to add that we know we all know that engineering is a trade-off right so and all of the approaches that we have for the hand nowadays for a dexterous hand there are pros and cons with almost all of them so the question that I always ask myself is that um yeah a lot of them works in short term but uh not many of them survive in the long term and uh the first company maybe to unlock this that a longlasting hand with uh that we can prove it after like I don't know 1 million cycles is still the same that that could be the the winning factor although I still believe one of the biggest advantages of tendon is just basically form factor because you can easily move all the actuators to elsewhere rather than being direct drive or being local but um yeah um it's interesting maybe we need to look back in a year or so and see how companies are deviating from their approaches >> I would like to add a Couple comments there, Meridat. One is that unfortunately I wish that uh you know a million cycle was uh high bar. Uh it doesn't if if you assume 247 operations, it doesn't buy you that much time. You still need you still need uh to replace some components and some tendons on a pretty regular basis. You can make different assumptions about the duty cycle and use cases but it still only lasts you for a few months and probably this is a trade-off that for some time we need to uh accept. The other thing that I think you are spot on about is the fact that the um tendonbased hand buys you volume buys you uh better form factor. I think that many of the companies that I see uh putting out tendon based hands uh are falling short of is they are using that um additional potential of having a better form factor to make a robot that has relatively slim forearms. Okay. So they're like, "Oh yeah, we can we can make it look like a human." And there is appetite for that for sure. But I think in especially in some industries especially in the industries that we are working with as foundation there's appetite for also using that potential towards sheer force like how how how how strong can I make uh this hand and accept something that looks a bit more like a robot from the cartoons in some sense like you know with this like big forearms but at the same time it doesn't have any limitation in tool usage, heavy heavyduty work, heavy duty object maneuvering while staying on the safe side of thermal performance of the forum actuators.

[00:17:08]>> I mean to your point the the the shopping uh bot um >> Oh yeah. Yeah. Iron.

[00:17:14]>> Yeah, the iron bot. Yeah, that's a case in point. Um and I I think uh they've just announced recently that they're ramping up production to 100 units a month by the end of this a thousand units per month by the end of this year.

[00:17:28]So uh with reason to but then that's a a limited use case for deployment and to your point like and and the point that me was saying was making was it's not just longevity it's also how much of stress um in addition to functionality uh that you can depend on for a long period of time because if serviceability is a problem um then I don't know it's a that's again another another additional problem for you know humanoid companies to solve right >> and I mean iron hasn't really shown their hand they had like a like a dummy hand on a with some circular actuators were going to the dip joints but they never really showed a working hand I think of their own design at least not a new one >> sure correct I think they showed the old hand on the new one but they have not shown the new hand except in in CGI so we don't know how it's doing it I I suspect It might be kind of like the Sharpa hand, direct drive at the joints is what they're trying to do. Uh so both both Sharpa and Wuji have direct drive hands, but they have a slightly different approach. You know, um Sharpa is uh directly at the joint, whereas Wuji puts it actually in the uh in the failins or you know, think of we're making the actuator at the actual finger bone and then they have to use a a worm gear kind of transmission to be able to get the rotation to work, right?

[00:18:57]Both those have drawbacks as Andrea has been talking about. One is heating for sure. The other is sometimes back drivability, but the big selling point of both of them is Cinderell is that it's a lot easier to get the cinder reel for that and that's why they are like the favorites of a lot of the different labs uh being able to do it.

[00:19:18]Um, but I think there's one downside that's going to be really hard for them to fight. And that is the fact that they just have so much inertia. When you put the actuators in your fingers, you just you can't get rid of that inertia. And that's the huge advantage that tended drive systems have. Um, but the downside that tended drive systems have is reliability of of course and the sim to real gap uh is pretty pronounced right now. So Andrea, your point about the forums not being necessarily having to be human shaped. I mean, why do you think all these companies are going for that? I understand iron and they want to have a very humanesque shape and for them that's a a selling point, but I mean a lot of the other American companies like Figure and Tesla, I don't really see the point in them directly going to as human design as possible to and then fight with having a slim forearm. Wouldn't it be better to kind of make it a bit more marvelous or like a big forearm to or would isn't the isn't the benefits enough to motivate a nonhuman form? You think >> my personal take is yes. My personal take is that you can either have the um robot look like a human or perform like a human because we haven't attained uh level of torque density that is comparable to what the human muscles have. Um in the rest of the robot we solve it by having less muscles, less degrees of freedom. In the hands we are trying to have a similar degrees of freedom as a human. And so you either make it weaker or you make it or you make it bigger. Uh which is why we at foundational are building that the forum the way that we are. Um but it seems like for a reason or for another which I don't know. Um also the companies in the west, the main companies in the west like Tesla and Figure have a lot of pressure on the industrial design side of things and and making the robot um look good, which in my opinion is going to be important at some point when you know like the industry matures and customers will have choice between uh a lot of different options. um then how the robot looks will certainly play a factor. I think that at this point we have an u exaggeration of uh emphasis on industrial design as opposed to functionality.

[00:21:53]>> I mean cases so far aren't really there.

[00:21:56]I mean the hand having a similar size and of course the function I think is important but it wouldn't really me as a surgeon if I had this size of a forearm I wouldn't be in that much of a worse situation I mean sure >> but do you need a bot to be a surgeon right now >> two and a half years now [laughter] three years this winter >> is that when you're planning to retire >> but but may I comment on this uh the risk form factor first of all based on what I've heard in the industry and what experience is that wrist is still a very much unsolved problem. Like there are so many so many issues within the wrist that if you really want to have a proper wrist with human form factor you have a few years to go. Um the second thing about the the volume of the wrist uh I kind of agree like we usually actually have a have a a watch for example on the wrist right so that makes it bigger than a human form factor. Some people just normally have a bigger wrist. Imagine there's a there's a person who is uh bulky and they they have no issue performing their tasks with a bigger wrist, right? And uh and lastly um it's a matter >> you're making fun of me.

[00:23:06]>> I've never seen your wrist actually, right? [laughter] Comparatively um but one thing that we know is that when as Andrea also address on industrial design, a part of the industrial design is basically the perception. So that people when they see something slim they correlate it they they connect it to that thing being more dextrous or being more capable or when it comes to the hand especially a more slim hand looks more capable uh no matter how beautiful or ugly it looks like h and a bulky hand on the other hand looks more lazy and less dextrous for example. So my thought is that the the reason they're putting so much emphasize on industrial design is basically for perception and that's where humanoid is today at least maybe in the US or maybe in the west that the perception of how capable is this system is is much more important that than how capable it actually is. I'm sorry to say that but maybe it's because of the VCs and investors but uh yeah that's that's at least my take on that side.

[00:24:15]I think that's very valid and um yes, unfortunately it is true. I think there will be a moment when the rubber hits the road and you know like you ask the robot to perform certain tasks and then if you're limited on strength, it's not a decision that you change overnight. Uh you're pretty locked in on an architecture that takes at least months uh to change. Uh one thing that I want to add uh is that I really agree with what Gustaf said, the shape of the hand itself is more important than the shape of the form and to an extent of the wrist. Just one note on what you said, Merduct, uh, which is yes, humans adapt to different shapes and including people with very large hands manage to manipulate very small objects. One factor that needs not to be discounted is the amount of brain power that we can throw at manipulation, which robots don't have yet. Uh, you can you can give me you can give me tweezers in in like both of my hands and have me go around all day only using those tweezers and not using my actual hands. And there will be a lot that I will be able to do with it. Perhaps not everything, but I'll be able to manipulate objects. I'll be able to eat. I know it because at some point I broke my hand and I like learned to do everything with a single hand. Um so that's an element that where two things have to converge. One is um I still think that in the short and medium term the hardware needs to converge as much as possible to uh the human shape uh primarily for training use and interface with the same environment use.

[00:26:13]But at the same time, if you told me, would you rather have the um hand be a bit more similar or the robot being twice as intelligent? I think that the second option would get us further.

[00:26:29]>> Very interesting.

[00:26:30]>> That's really interesting. That reminds me um of what we've seen clearly from the demos of Figure and the Helix AI stack where they started off with the S1 and the S2 layer, but then they got the they added the S0 layer for more granular uh motor control.

[00:26:46]And that seemed to make a marked difference, right?

[00:26:49]>> Yeah. And the same hand being used >> the same hand. Yeah.

[00:26:52]>> Yeah. So talking about the hand and it being similar to us, do you do you buy this ID there? the the post of it being the most advanced robotic hand yet. I mean, the more I look at it, the more it looks like the the bot hands I saw four or five years ago as free downloadable SDLs basically. I mean, the wrist design I really like is similar to you. It's a fancy way of solving parts of the problem. But when it comes from like the palm and out, it has a very similar design to a lot of different pot hands out there. And [clears throat] I mean, is there anything else in this hand that you see that you like, "Oh, this is a really nice new ID or a variation that I haven't seen before."

[00:27:34]>> My take is that you judge how good a hand is from performance. And so far, the most impressive thing that I've seen out there is the demo that Genesis AI put out using the Vuji hand. Um, I think like I said that in the medium to long term, the Vuji hand has some limitations, but um I that's still like the most impressive demo that I have seen so far. Nothing that I've seen here is extremely uh impressive. Um, there are some very original and interesting architecture choice. There are things that as an engineer I am very interested to understanding more of. For example, their multi-layer tendon stack I think is very interesting. Um how the uh you know birectional um tendon actuation works is interesting. What are the pros? What are the cons? But um I wouldn't say that this is the best hand that I've ever seen. Yeah, I mean I agree in a way. I like that there are new versions showing up, at least variations of them, but I still believe that having a few actuators within the hand will help you in many ways. It might be too much of an issue when it comes to the heating and so on, but I'm still looking forward to like a hybrid design with a few in the palm and the rest in the forearm similar to the way that we were made up. But I mean the the good thing with this design it's very hollow so they have room for all the tendons. I mean if you put in a lot of jackers or others perhaps you wouldn't be able to have as many tendons passing through but what you said in your post earlier I I never really thought of that in being but yeah we never see them actually doing this motion. Um so when it comes to anatomy or human how we move our hands different sometimes they mix up the words and so on but the the movements that we need to do is this one that we go put it to the side of the finger and then out and then of course putting it into position for opposition. So the the flexion and then this is often also called adduction and abduction that way. But I mean they they only show the rotation in a way and then the flexion. So maybe that's enough. But we do a lot of things where we push our thumb flat against the side of our fingers and having a flexion there might be okay. But then you would get your pinser grip which they I guess use for a lot of things. And that is not ideal when it comes to certain tasks like turning a key. This is a lot more tricky way of rotating a key than having a flat thumb for instance and other things. Um, so I wonder if it's just that it's hidden there as you say or one of the posters commented that yeah, there is an app adduction down there hidden next to the rotation and the flexion and extension. Um, but similar to that, we never really saw this being actuated by itself aside from it moving when it flexed its fingers. It kind of moved a little bit, but I wonder if they they just skip having that actuated and having it as like a an effect of the finger flexion. It will just follow along.

[00:30:53]>> Yeah, I don't know. There are there are a lot of things that if you look at the hand statically you assume it would be able to do and then doing the motions it doesn't do. Um I have again like two theories. The one that I mentioned in my post is that for example going back to PIP and DIP my take in the post is that they were actuated with a single tendon. I'm not sure about that. uh someone will message me claiming that that is not the case.

[00:31:24]Another theory that I have is that this architecture that they have of the closed loop where every joint is actuated with a birectional uh actuator and a loop of tendons that goes makes it to the joint and comes back um might create problems for um for controllability. What I mean is that for example, if my um if I have a tendon loop that is making it here to to MCP, right? And then I have another tendon loop that is making it to PIP and then I have a third tenant loop that is making to DIP, right? The tendon loop only works if the um total length of the loop stays constant. But this is true for MCP. As I move this, like this remains constant.

[00:32:16]If if I move this in this fashion, the length remains constant. But the length of the second loop between this configuration and this configuration changes and the length of the third loop that makes it here between this configuration or this configuration changes drastically and that makes the IP harder to control.

[00:32:38]>> I'm sorry. Um, isn't it the same case about wrist like wrist moving and then affecting the fingers? But they solved that. That is the case in many um almost all >> in many in in many tenant based hands.

[00:32:52]Correct. So and that's what prevents most hands from using birectional oxator. That's why Kyberlab doesn't have a wrist. Like it's the you know elephant in the room is that they always show the hand but they never show the wrist.

[00:33:06]>> Uh they solve it by having bowden tubes that go directly into the palm. And the bowden tube has a fixed length. It works like a a bicycle brake tube. So in your bicycle you have like the bottom tube going along and the bottom tube what it does is it it creates a compression path through the spring that composes the bottom tube and uh in that way the length is is fixed. So up until here they are able to up until the knuckles they are able to manage uh the path length but then when they go into the fingers they don't have space for the bowen tubes and um and so that becomes hard. Another note on the bowen tubes is that many companies complain that they create a lot of friction which they do.

[00:33:55]They increase friction which decreases um tendon life and decreases controllability. Um the it seems like uh 1x spent a lot of engineering time into the full tendon stack. The reason I say that is because if you see at this picture or like at any of the frame so the little two ropes that you see coming out of the center of the actuator are act are the actual resistant material.

[00:34:25]uh might be might be dene but some stiff rope and that's very thin compared to the size of the of the uh of the bottom tube. So I assume that they have at least one intermediary layer between the rope and the tube which is a sacrificial layer uh that takes care of the friction and maybe consumes over time. I think Scott had a take Scott had a take on what on what that is, whether it's daea or something else. Scott.

[00:34:56]>> Oh, yeah. I'm suspecting it's probably also da and it's just that the uh the internal spring structure seems to be a square cross-section, which means you would already have a lot of exposed edges in there, which is probably why you would have also another like tube inside the tube, probably made out of teflon or some sort of coating so that when the tendon is rubbing up against uh that coiled spring, it it is not fraying. So, I kind of agree with Andrea. There there's probably some sort of sacrificial layer in there which might be, you know, a tube inside which I think some high-end bicycles uh you will see that as well.

[00:35:31]You will you have the cable itself, a tube around the cable, a spring around that tube, and then an outer tube around the spring to hold the spring together.

[00:35:39]So, there there's a lot of components that have to come together. The advantage that bicycles have is that they don't have tight bending radius radi anywhere and as a result they um they can they can afford metal tendons um which are much more resisting in in a bowden tube than than a polymer tendon is. So, do you do you want to spend a little more time on your on your third on the third part of your post, Andrea, about the the wrist?

[00:36:14]>> Yeah, absolutely. So, like I said, and there's there are other uh All right, if you can open this video. Yeah, we created this uh this prototype uh with my team here. It took essentially it took the time it takes for the dynamic salts which are the actuators we're using at the bottom to arrive to here and then to 3D print. But this is how the wrist works. So you can see Max here uh my team uh moving the uh two bottom blue wheels. One of them is hidden either concurrently or in opposing directions. Yeah. And you can see that that generates u motion along the two degrees of freedom of the wrist. And obviously >> I saw this video from you. It was it's ingenious. But do you think this is >> I don't know exactly.

[00:37:02]>> Talk about the life.

[00:37:04]>> Yeah. How durable is this?

[00:37:07]>> It Well, basically it's the Optimus wrist with cables. It's the same idea.

[00:37:12]>> Is it superior though?

[00:37:13]>> Which is a big difference in my opinion.

[00:37:15]Yes. Yes.

[00:37:16]>> But but actually it might actually I might take that back for a second. Um, is the order of operation yaw, pitch with their wrist? Because I think with the Optimus wrist, it's usually pitch yaw. Correct. On the card joint. So, I think the the arrangement of the joint, you know, well, it's kind of a similar idea of having the differential loading on there. This case being done with cables rather than a linear drive. For some reason, they think they decided to do the ritz yawing first and the pitch afterwards. And I wonder if that has something to do to help accommodate the the the bow and tubes a little bit better, just the order of operation.

[00:37:53]What do you think?

[00:37:54]>> I can't comment on Optimus, but I'll tell you what the differences that I see are from uh linkage base. You can do it either way. You can do it with the joints in one direction or another, whether you are using linkages or um or or tendons. Um I think the advantage here which you don't get with linkages is the ease with which you can adjust the um transmission r transmission ratio or gear ratio between the the the movement that you have at the motor. So the motor on the bottom and the movement that you have of the wrist. So you know how in a bike depending on the gear that you take the size of the uh wheel that the of the gear that your chain wraps around varies. This is a similar concept. If you go with a very small wheel at the bottom on the motor and the very large wheel at the joints then you get a wrist that is slow but very strong. If you use like given that you use certain actuators, if you do the opposite, so if you have a very large wheel on the um on the actuator and smaller wheels on the joints, you get a faster but weaker uh wrist. So what this gives you that a linkage base wrist doesn't give you is the possibility to um adjust that gear ratio and trade off speed for power which is something that you cannot do uh as easily with linkages or if you do >> yeah I was just wondering can't you do the same with like a linkage base is basically a four bar mechanism so you can change the length of one of the bars right >> correct the problem with that is that first of all you don't have a constant transmission in a in a four bar linkages. Uh the moment arm of the uh link about the um end joint changes throughout the range of motion. Um and you can change the size of the of the of the bars but at the expense of using more or less volume. What I really like about this design is the compactness because given a volume like that's all that the wrist is going to take. The tendons slide but they don't uh you know they don't need volume for range.

[00:40:30]>> Yeah, I I think that you make some good points there Andrea. One is that the uh the torque is linear throughout the entire range of motion unlike with a linkage mechanism. The other problem with a linkage mechanism is that the links get in each other's way. So the range of motion is usually limited in in the different directions especially like the yaw. So I think with this you can probably get greater range of motion both uh in the yaw and in the pitch uh which are you know one of the advantages of doing something like this.

[00:40:58]>> Yeah. My the main advantage in in in my opinion it's my it's my favorite thing of of this of this design. The downside to what to the point that Gustav was making earlier is on um you know necessity for periodic maintenance to tighten the ropes potentially. Um, I like using a closed loop rope here better than I like it on the fingers because uh the bending radi are uh much larger and you can afford to use a thicker rope which you can secure um in a in a like more easily um as opposed to having to maintain these very thin ropes and making sure they don't creep and they don't slip. But um at the stage of the industry right now any tenant based uh hand wrist or whichever it is will require uh periodic maintenance.

[00:41:58]Unfortunately um this is easier to do when the application is an industrial robot or uh you know a robot that lives in a certain workspace. Much harder to do when you try to deploy robots at home.

[00:42:12]>> Yeah, that's true.

[00:42:15]Method, uh, you shared something interesting with me. Do you want me to bring it up now?

[00:42:19]>> Yeah, please. Uh, I saw it recently actually and it was I'm sure Scott would love it at least. I don't know if they've seen it before, but it's one of those very genius mechanisms that you see once in a while. And uh, >> so this is from Barto Rajivski. Am I murdering his name, butchering his name?

[00:42:35]>> Founder of Bionic.

[00:42:37]>> Mhm. And I was just simply thinking okay as a passive system it works well but if you want to make it active also you can basically have small brushes motors underneath and uh it gives a very interesting um uh direction. the the only problem is that you have that this constant contact in the middle which you probably get a lot of friction on that ball joints but uh otherwise um yeah I thought that these are this these could be type of like alternatives that one could use for uh wrists >> well isn't that a point of failure an easy point of failure >> 100% and also the fact that the middle of the mechanism is filled because we usually need to for example tendon drive we need to route the tendons probably from the middle but in this scenario we always have to have the midpoints filled basically >> yeah the the design would would have to be a bit different to make it a risk the main downside that I see with it is that it's inherently uh limited in range of motion what we are seeing here is roughly um maybe 15 20 degrees of like in each direction might be enough for um for radial and ular deviation definitely not enough for flexion extension.

[00:43:58]>> Um but there are a lot of of very interesting mechanisms in the literature that are not necessarily similar to this but have similar uh ball jointed type of um of movement.

[00:44:13]>> Yeah, these mechanisms you can almost never achieve 90 degrees on any any direction. I mean previously I know he left now so I can say it but Scott previously said that do we really need 90° in the wrist I can't do what he says but I mean we really can I might be a bit more flexible but but especially when you put like force against it we we can really go 90 degrees and further most of us unless you had a sprain or something. So I think 90° is a good thing to have but also the radial deviation is of course not as big but you need a little bit there to for handling objects and the instruments.

[00:44:52]>> All right so uh moving on we've got an interesting post from uh Scott Walter and Andrea. Do you want to kind of chime in on what's going on because Scott is confused. Why is Neo's bottom half a unitry G1?

[00:45:07]>> Scott, have you ever heard of generative AI? It's >> quite these days.

[00:45:13]>> Yes, I have. Yes. [laughter] Yes, I have. But I I just I thought it was kind of funny that to to see uh that obviously generative AI has is seeing more examples of the unitry running around than anyone else's bots. So when you say put together a robot, a humanoid, that's what it's thinking right now. And a few years ago, it was kind of when you asked it to do something, everything looked like Optimus.

[00:45:39]>> Yeah.

[00:45:40]>> And and now it's like it's looking like something else. And and I think just the irony is here is that it's reposted um by the the you know official head of marketing for 1X. It's like I mean, wouldn't you have looked at it at first?

[00:45:54]Kind of embarrassing. It's not a good look. And it's also kind of the Winnie the Pooh look, you know. [laughter] It's like only only running around with a shirt.

[00:46:04][laughter] >> Yeah.

[00:46:08]>> It's a pretty good video though. It's quite funny with the whole cap and and >> Yeah.

[00:46:14]>> thing under.

[00:46:15]>> Yeah. No, it's it's just sort of funny.

[00:46:17]You think that when you want to control your own content, you'd make sure the content is actually your content and not competitor.

[00:46:24]>> Yeah. I I I mean I look at AI stuff all the time, but I actually didn't consider the fact that it wasn't their bots just using a unitary body for >> good. But yeah, for for those of you wondering, he's talking about Dar's post. Yeah.

[00:46:38]>> Yeah.

[00:46:39]>> Where this video was. Yeah. So, >> and this was this was like a fan video made by someone else. So, it wasn't made by him. And then it's it's interesting. And and of course, the funny thing is is the hand that's on there is the old Neo hand.

[00:46:53]It's not the new one.

[00:46:54]>> So, I mean, there was just so many levels of [laughter] of things that were kind of wrong with the video.

[00:46:59]>> And and this part about the phone, I'm not sure if none of the hands, humanoid hands today, they can interact with the phone. They don't really have a stylus on the um >> fingertips.

[00:47:12]>> There was one video actually from 1X, I don't know if you remember a few months ago, but somebody asked Neo in a show or something, can you take a photo of me with my own phone? And Neo said, "I cannot take a photo, but if you if you put it on a video, I can just record a video." It means that the finger cannot actually interact with the with the screen because I had actually this issue. I wanted to make a hand that it can or just a simple gripper that can interact with the phone and it wasn't that easy.

[00:47:41]Actually, you need proper connection to ground even if it's a stylus.

[00:47:46]>> H interesting. didn't know that.

[00:47:51]Surprising, right?

[00:47:52]>> Yeah, it's it's one of those. The second thing is that almost none of these humanoid uh hands they have shown um handwriting something with a pen or a marker.

[00:48:04]>> Uh because that's where the accuracy becomes very important. I don't know if have you seen any >> I think I've seen someone attempt to do something but nothing great.

[00:48:15]>> Yeah.

[00:48:16]did with with uh with the um how's it called? It's not Dex hand u robotics. Yeah.

[00:48:25]>> Yeah. They had but but they had they had a fixture I think on their fingertips and they were showing that they could uh draw circles for that.

[00:48:33]>> Yeah. But that's a circle. That's not calligraphy which is I think something way more >> complex. I mean that's extremely hard.

[00:48:40]Yeah.

[00:48:40]>> Yeah. Yeah, >> it has more to do honestly it has more to do with uh fine motor control of the whole actuation chain than uh than movement of the hand itself like you could in well >> Mhm. I agree.

[00:48:58]>> Yeah.

[00:49:00]>> Use these muscles when we write with a pen. The wrist is stiff and of course we move the whole package to the side.

[00:49:06]Yeah, writing is in this area of the hand. Whereas drawing on a whiteboard, then you then you fix that this and you use the wrist and the elbow and shoulder rather when you draw on the whiteboard.

[00:49:18]So a whiteboard, these bots, all of them could probably do pretty well. Whereas writing on a piece of paper that I don't think any of them can do so far.

[00:49:26]>> Yeah.

[00:49:27]>> Aside from like making bigger squiggles, which is actually drawing with their >> Yeah. it it becomes kind of the you know the arm and hand equivalent of whole body control is that normally in robotics we we literally cut them off right here and we solve what's going on with the hand and we or and then the the arm separately as completely separate problems one position the hand make the hand do whatever [laughter] reposition it but the whole idea of integrating that is a is a much bigger challenge and that it it's not just a software problem but I think it's also a mechanics problem is that you really have to know how the two interact and work together and so eventually I I think you will see that the ARM is really comes all the way out to the fingertips and you're not seeing it as a separate problem.

[00:50:11]>> But then as we talked previously about the whole AI stack or the compute having the bot walking and taking notes at the same time that's going to be >> and chewing gum. Yeah. Chewing gum.

[00:50:23]>> Yeah.

[00:50:24]>> The reason I brought that example is that most of the task that we see in human demos a six years old can do. But a six years old cannot necessarily write something properly or a lot of other examples like I haven't really seen many tasks from humanoids or dextrous hand cases that five six years old cannot do.

[00:50:45]>> So if I do you need a bot hand that can write >> I mean the writing is not really the point. I think the point is imagine if it's if it's going to be a chef for example and it's going to cut sushies or like salmons, right? Like those kind of tasks, they all require a level of control that I still haven't seen in almost none of Yeah.

[00:51:09]>> Yeah. And and and that's basically a if it can do handwriting, that's a proxy for a lot of the other things you'd like to see it do. Exactly. So I think it's a fair test.

[00:51:18]>> Murd has the perfect business ID.

[00:51:20]Instead of buying a lot of humanoid bots, we just hire a bunch of 5-year-olds to work in the [laughter] factories.

[00:51:28]>> This this was a thing. This was they forbid it. They forbid long ago.

[00:51:34]>> Actually, I I had a comment a while ago, one of the sessions. Yeah, it was about that. I was saying I was trying to defend a smaller form factor and I said actually in in the Chinese factories they're using very small people because they have a smaller hands and it goes into the PCB and details. So um yeah but yeah I hope that >> you know we're going to get a lot of hate in the comments [laughter] >> walking a fine line.

[00:52:03]Let me steer this wonderful discussion towards the next non-controversial topic.

[00:52:10]>> All right, so here we go.

[00:52:12]>> Our very own Meredith Farimani has been slaving away. He put up this really interesting post on LinkedIn. He said, "We need more data, quote unquote, has been the common denominator in most of the conferences and humanoid robotics discussions I've attended over the past year." But I also felt that the common understanding of how this data is collected and what the trade-offs are is still not really there. That included me. And then you wrote this the state of data collection for robotic manipulation. Cthosis in many ways take us through your thoughts.

[00:52:48]>> Of course, of course. Thank you very much for bringing it up. Uh so to be honest the reason I wrote it was for also myself to learn more because it's uh it's only one and a half two years that I've got to uh got myself famili familiar real famili got myself acquainted to uh data collection uh when it when it comes to especially humanoids and it's a it's a fairly new field actually. So the background is for the audience who are less um um aware of this matter uh the background is that uh we had LLMs fairly quickly and they're working very well because we had so much text data on the internet and then we had the VLA VLMs and multimodel u models also because we had a lot of video and photos over the internet but when it comes to robotics text and video are not enough sources of data. So we need uh we need so many other things such as inertia of friction uh forces and all these small small data points that uh that we can enable a robot to interact with the with the environment. So what this scale is trying to say by the way this is not accurate. This is some something just approximate is trying to say that what is important in data collection for robot manipulation is is this method scalable. That's the first question. And the second question is how much fidelity here means how easy it is to convert this collected data into something useful for the robots. If you want to put an example here, we would say direct tele operation is not really scalable. For for many years, the the industry was focusing so much on direct tele operation and then as a result maybe or maybe not train the robot over that. That is not really a scalable solution. But at the same time, synthetic data over simulation is very scalable. But these two also sits in other ends. So direct operation you can uh use it very very much directly for the robot but synthetic data or for example egocentric data which is very popular nowadays uh they cannot really be translated directly to the user.

[00:55:03]There's another problem with egocentric uh human video for example which nowadays wherever conference you go there are a lot of people actually coming to you and saying oh we are uh collecting data from uh workers doing sewing or uh doing uh I don't know uh manufacturing work and what they do basically is that they put a camera on the forearm uh forehead sorry of the of the users or of the operators and they do a job and the camera is trying to detect ffects how the hands are moving and how the job is being done. In these scenarios, we know that a lot of data are missing, but it is still useful in some scenarios, but in in like proper training of a robot, they're not really useful enough. So then we go some middle um solutions such as uh Yumi or UMI uh solution uh like a portable proxy device. This is very very popular nowadays. So it's basically a handheld device. Sunday robotic is a good example, right? So a a handheld device that is very similar to the grier of the robot and then you try to do the task with that one and then it translates to that certain robots directly. The issue with this methods is that is that um then it's specifically made for that specific robot. If you want to scale it or expand it to other robots or to other scenarios, it becomes extremely hard.

[00:56:26]And also we know that two-finger grippers or even three-finger they are still falling short in some of the applications some of the very dextrous applications. Then we have the gloves that during this uh podcast we have talked a lot about different type of gloves uh including manus gloves and um they are also very good uh solutions.

[00:56:46]Then the problem is another axis which is cost. So some solutions are extremely costly, some solutions are extremely cheap. Like egocentric is extremely cheap or synthetic data is extremely cheap, but they're not as effective as a glove. And so you see there are a lot of pros and cons with all these different solutions that we have in robotics data collection. And I try to basically just explain all of these um shortly in a simple language, including their pros and cons. Have you have you considered the bionic hand in this >> uh bionic hand as a data collection device or as a >> like like psionic >> like psionic?

[00:57:25]>> Very good point. Very good point. I haven't considered that because that that was very unique application that um um he said just to repeat it for other people. So psionic hand is also a prosthetic hand and the when the people who are let's say amputes or disabled so when they are working with that hand that data is being collected from that hand and being used for the robot hand which is very pretty much identical to the prosthetic hand. That is a very very unique approach. The reason I didn't really include it here was that I couldn't really find so much resources about how that process is actually being done. I could interview a deal uh on that but uh and I'm not sure if that is a very scalable solution to be honest because not every company have an identical prosthetic hand and we don't probably have enough uh people to to use it.

[00:58:19]>> No. Well, uh I'm hoping to get Javanni Zap Zapur Zapur. Um Andrea, am I pronouncing his name right? from Bionit >> because they >> Okay, close enough. I'll survive. I'll take that as a win. But um Javanni has also been working on the bionit hand which is again a lot like the psionic hand but more it's it's it's um their focus is more on uh robot training rather than prosthetics because um I think what Adil is is doing with psionic is he started off with it as a prosthetic and then you know the robotic training applications come later. So I think that's going to be interesting.

[00:59:06]>> Correct. I have a question.

[00:59:09]>> I don't think sorry just quick take. I don't think that those two industries are converging. I think that they are diverging.

[00:59:16]>> I think that in order to have a good uh a good uh prosthetic hand. It has to be very lightweight. uh to my point before you can throw a lot of brain power at it and even with simple motions.

[00:59:31]Ideally they come from the uh from the amputated uh limb but you can you can ideally like command it and a very simple motion is may may be sufficient.

[00:59:42]Whereas with robotics we are seeing the direction seems to be uh you know like more actuator heavier more degrees of freedom and and it needs a lot of energy to be operated. So those like prosthetics and robotic hand look the same to an untrained eye but when you look into the incentives of what the architecture needs to be they are very different.

[01:00:09]>> Very correct. I want the hand surgeon to weigh in here.

[01:00:14]>> Say again.

[01:00:15]>> Gustav, the hand surgeon, will you please weigh in?

[01:00:19]>> I agree. My hope is that it will feed into each other in a way that the prosthetic side gets benefits from the development of robotic hands. And I still think they can to some extent with making better actuators, shrinking them so we could have better prosthetics. But I I agree. I mean, you're not going to put the whole CPU into a small forearm of an ampute, which probably would start down here, your prosthetic, since you want usually have a small stump to connect it to. So, the the the size is very limited compared to what you could put in your bot. Um, and as you say, will be like edge compute and today at least not yet. You can't make it to too long, I guess. Yeah, but still I have some hopes that it will at least parts of it will translate but u we'll see. I think I think it will especially the software stack and controllability stack but it will be a byproduct of if the volume of the robotics industry is what we expect to be uh then like fortunately the the >> seriously >> I think I think I think it's going to be I think the I think the uh humanoid robots market will be enormous. Yes. Um and I think that it will bring some benefits to prosthetics. Hopefully the prosthetics market stays small and ideally shrinks over time as opposed you know like you Yes. Um but to Gustav's point I think that the like there will be some cross function like cross benefits cross domain benefits um but it will not be the same product. I'm I'm pretty certain that it won't be the same piece of hardware.

[01:02:05]Uh another thing I can add actually is that um I don't know how Adil or anyone who is taking that approach is solving the issue of vision because we have the hand and we have the force sensor and torque uh force sensing and also uh friction and everything but we don't have the vision yet. So we don't really see what is going on there. So we are collecting a bunch of data based on special uh special um >> and also you need you need the vision to deploy the hand where you need it to be used as a tool.

[01:02:39]>> Yeah. In the end we we need the both the vision and the tactile kind of coming together >> tactile. Yes. For for if you're reaching around a corner out of line of sight.

[01:02:48]>> Exactly.

[01:02:49]>> Yeah. So that's >> what's the number what what's the when you dot plot it what is the number of kind of functions that fall under these two categories.

[01:03:02]>> Yes. So I have a list actually that um so there are up to let's say 15 different data points that we need to collect roughly but the point is the way is not the same. Some data points are much more important than the others. For example, one very interesting data collection point or data point for robots is sound. And I'm not sure if many people have worked on it. So sound basically means yes, imagine you have this bottle and in a manufacturing line and the robot wants to close it. So as a human when we close it, we hear this click sound, right? So we know that the task is done. It's properly closed. Uh but uh as a as a robot you are missing the sound or voice basically uh or noises.

[01:03:49]>> What about vibration?

[01:03:50]>> Well, vibration could be a point but it's not directly translating to the same. There could be a lot of other examples. Another example is imagine a robot. So when you when you actually put your mug or glass under the water so the water is falling on it and the sound of the water is changing slightly. It's not Doppler effect. I don't know what is it called actually. But the sound of the water pouring inside the glass is changing and that's how >> the column of air is reducing.

[01:04:16]>> Yeah. Yeah. Yeah. The resonance chamber is is changing in size. So it it goes up in pitch.

[01:04:22]>> Exactly. So imagine when you're filling your glass of water in the dark at night, that's how you realize how much is being filled. So sound is actually a very important data point for >> also textulation. [clears throat] >> When the box squeezes your hand too hard, you scream and then it >> [laughter] >> Yeah, >> that's an extreme example.

[01:04:43]>> You hear the sound of your bones crack.

[01:04:45]>> Yeah. [laughter] >> So guys, I have to apologize, but I need to drop off. I look forward to listening to the rest later on.

[01:04:52]>> Thank you for joining us. Truly appreciate it.

[01:04:54]>> Welcome.

[01:04:56]>> Um, I want to quickly move on to Scott.

[01:04:58]This was a really spicy take from you.

[01:05:01]And let us talk about this now.

[01:05:05]>> Boy oh boy, you set the cat amongst the pigeons, didn't you? I I don't really see why it was so spicy because the the reality is just about every roboticist I've talked to over the past several weeks and months and even after this went out all said that's pretty much what the state is. Um we we have to like I think some people get a little bit confused what do they mean by a prototype. Yeah, there are a lot of people who can throw together a very simple robot prototype and say look I got something here. But that's not what we're talking about. We're talking about something that's fully functional. So in the the normal context of prototypes, we think of like an automobile. Everyone was talking about an EV that was like really easy to kind of put together a something that resembled an EV that was like pretty good. For instance, when the Cybertruck came out, I would say that was a solid prototype. You could drive it. It probably had everything that you needed. The thing is they weren't ready to scale it to massproduce it. And that's really where the the discussion was like the production side of that is very hard. It's really easy to to get together a pretty good prototype of an EV that you would almost say is production grade just that you know you have to now do the hard work of figuring out and in a sense that hard work is rolling up your your shirt sleeves and just getting out there you know having to build the factory and get the automation getting all that stuff there's a lot of engineering that goes into getting getting that part right and I have joked all along that that when Optimus came out I said this is going to be the first time where it the the prototype is going to be hard >> and the production may be relatively easy.

[01:06:40]>> Is that a good thing?

[01:06:42]>> Well, um this is what I think everyone has to understand is no one has a prototype right now which is worth scaling.

[01:06:51]>> Not even figure >> no I mean even even Brett admitted you know he he said that they're they're doing figure four for that reason that that's the one that they want to start scaling up in larger quantities. that figure three right now is looking to have a scale that's going to be below 10,000.

[01:07:08]And now you do need to scale a little bit in your prototyping process because we know there's this problem with SIM to real that at some point you've got to get enough bots to get them out there.

[01:07:17]And not just from the training aspect, but just to find out like reliability.

[01:07:21]It's like how long can we run these things before they really start breaking because theoretically we might think it's robust enough, but then we find out all sorts of things. So you need a lot of real world data. So you need to get a certain amount of scale. Now unitry definitely has demand for their bots. I mean if you try to buy one of their unitry bots right now you have to wait a couple of months. So um so clearly they have demand and they have a little bit of supply problem. So, so why aren't they ramping up? Because their robot isn't really worth scaling past 10,000 per units per year, which is about what they can do with the current production techniques, which is billets that they are just machining because they have a plethora of CNC machines that they can just use. But you run up to a limit there. You can't go much beyond that without saying, "Okay, we want to scale it." Now, would it be hard to scale?

[01:08:07]Well, in theory, it wouldn't be that hard because they should just start making these exotic castings and a lot of the other things. They they could get that up really quick. It's just that that takes time, right, to go out and start putting the the the effort and the and the design and the tooling and stuff like that that now you can start producing quantities that are in the hundreds of thousands and maybe even get close to a million if you really wanted to be able to do that. They're not doing it for a reason because they know that the the bot is going to be obsolete maybe in six months with a much better version you can have which still might be not quite good enough there. And there's a lot of the reasons these prototypes are not ready. One is the hardware still is not robust enough.

[01:08:46]There's there's a lot of problems with that and the software isn't ready either. So while they can do a lot of things, it doesn't quite have that full capability that everyone expects to have from a robot that's that's going to be some that's going to be useful for scaling. And then part of it's a chicken in the egg. It's like the reason why the software is not good is that we haven't been able to get the training data because we don't have enough robots. So you get this kind of flywheel that you're going to have to go through that at this point, yeah, Tesla internally may have a demand of thousands of robots just for their training purposes. they might want to scale up. But from all the indications and like everything that Elon is saying and like don't listen to me actually listen to what he said not only several times but most recently that is in this same month that we're in right now and that it was actually um an unprompted reply to someone else's post who was very enthusiastic about the Optimus ramp up and made it sound like they were going to be large numbers. And there's also been a couple other posts claiming the supply chain about a 100,000 and everything else. It's like that's not really what I am hearing and seeing on the supply chain and the fact that it's it's going to be very limited.

[01:09:55]And then if you listen to Elon himself, that's the source. That's the footnote right now. Production will be extremely extremely not slow, extremely slow.

[01:10:04]Okay. Now, I'm assuming that Elon, like everyone else, when they say extremely slow means like really small numbers and not like he thinks a million is a small number. Okay? I don't think it's like that. I think it he's he's really using the same vernacular that any one of us would be. And that is, you know, it it might be that they're only producing a couple of bots a day when they start going in and they slowly ramp up and they slowly ramp up and slowly, but you're not going to be seeing a thousand bots coming off the the assembly line in September per day. that's going to take a lot longer to be able to do that because they are talking about the fact that they've got to get everything up and running. Lars also listened very closely to his statements that was uh on on Herbert's channel about that that he was being very careful about what he was saying and I think it was a bit misinterpreted that sound like oh you know they've they've got one Optimus production line up and going and there's going to be 40 of them and it's like no no no they've got one subline of the 40 that are needed to put it together which has gone through the field.

[01:11:03]>> That's the modularity you was talking about. Right.

[01:11:05]>> Right. Right. Right. Right. And and then also look at the the image. You know, Tesla showed us how the SNX line was torn down. So they shown up being torn down. Now it's a big empty hall waiting to get filled up with equipment. The picture that Elon posted. There was no production equipment there. And anyone says, "Oh, I see production equipment there." It's like, "Okay, the production equipment you're seeing are behind them." And that is the Model 3 and the Model Y line. Okay, that's the existing production line. But the Optimus line there being all you're seeing is concrete floors. So what you know every what I'm trying to do is is temper expectations which is is exactly what Elon is doing and he's done in multiple statements and if you go back to his statements I think the end of the first quarter when he was directly asked you know how many optimi do you think you'll be uh ramped up to the end of the year and he says we have no idea because it's very difficult to know he laid everything out he made it very clear and he said the production is going to be very very slow and again unsolic ited you came right on out and and said when he was seeing that okay everyone's getting a little bit ahead of the skis here getting a little bit excited about it it's there so that's what I'm trying to remind everyone is that production uh not only is hard but so is a prototype >> and in many cases they're pro they they they're because remember what Andre we've been talking about the hands today right that's the most important part of the bot >> so as we can see >> we can get dancing bots and cartwheel bots and everything else, but what's really hard is to find a bot that has a hand with the dexterity that can actually do a lot of tasks as well as seeing long horizon tasks. So there's still a lot of things that have to be done before you have a bot that's really ready to go.

[01:12:50]>> I may add something here >> please.

[01:12:54]>> Yes, thank you. So uh what I wanted to say if I wanted to rewrite this sentence you said so I would say prototype in in humanoids is extremely hard and as for the the production we don't know how hard it is.

[01:13:06]>> Yeah probably it's so ambiguous like no one has ever done it properly >> for us to know for for automotive it took over 40 to 50 years until they figured out the production line and be able to streamline it. Right.

[01:13:20]>> Yeah. Yeah. And and the thing to point out is is some people think I'm saying, "Oh, you're saying the production is easy." It's like, "No, no, no, no, no.

[01:13:26][laughter] I didn't I wasn't changing the words around this. It's like, you know, it's really prototype is hard and production will be hard." So, leaving that aside, you know, production, >> I agree. We we don't know. But before we can even get to production, we've got to get a prototype that is worth building.

[01:13:42]So you think people are really still stuck in this, oh, we need to get something out there to collect data so Tesla will be okay with releasing a half >> Yeah, I mean that's that's that's a whole that's the whole thing is you get this chicken and egg that's going on here is that they they've got a prototype of some sort. We're not we have no idea where the state of it is, whether it's got full capability or not.

[01:14:04]something they're they're ready to kind of build up, but um in order for it to be really performative because we need to see something that is very very performative, they need to have the training data. So they may be doing a lot of sim to real right now, but at some point they need to have the real bots. So you really won't be able to show it doing anything except maybe walking around and and no one is going to be excited about that. So again, the bar right now for a humanoid demo is very high. It's not like what was AI day when Bumblebee just kind of or Yeah.

[01:14:36]Bumblebee came out and stumbled around on the on the stage and everyone was very excited about seeing it because no one had ever seen anything like that.

[01:14:42]Now we're seeing the cartwheels, we're seeing them running half marathons, we're seeing all these other things. The standard is very very very high for Tesla at this point. And there's I doubt that right now the prototype is at that level and it's going to take time. And part of that, yes, is getting to production so that they're not having a couple of handbuilt ones, but they get enough that you can start to get that flywheel that maybe you have a room full of a couple of thousand and then you get the data and then maybe and that's why, you know, it just it this is my opinion that it's we're probably not going to see the unveil of V3 until next year. You know, it was I was hopeful that maybe it would be late third quarter, fourth quarter, but you know, more the signs are like it's probably going to be a little bit later.

[01:15:26]And of course, that is not the one they're going to scale to a million units.

[01:15:30]>> And and V4 is not the one. Probably V5.

[01:15:33]Again, >> what has Elon been saying every year we're going to come out with a new version.

[01:15:38]>> Yeah. So to that point, Scott, um again I come back to the question of of lowhanging fruit and useful work that can be done immediately with whatever there is and what is your thought about what scaling means if V3 were to be scaled as is with what we've seen and what we think it might be capable of.

[01:16:04]Well, the thing is you you don't scale a crappy bot. I've said that before. So, at some point, but you do have to build enough of them to find out what are the issues with it that we have to improve.

[01:16:15]So, >> is there another word you use instead of scaling because I think people are getting kind of kind of this whole debate about >> maybe ramping ramp ramping ramping up a little bit. So, so the thing is before you start talking about a million bots, you've got to start talking about a h 100red thousand bots. And before you talk about 100 thousand, you better start talking about 10,000. And before you start, you better start talking about a thousand. We're barely talking about a thousand for most of these companies right now.

[01:16:41]>> A few of them have come even close. And the ones that have hit 10,000 are really putting together with bots that are mostly disposable. You know, it's like they they're they're not adults size.

[01:16:51]They're, you know, they they overheat.

[01:16:53]They can't do a full-blown payload. They don't even have hands. Um, so they can't do any manipulation. Those are the ones that we're seeing being scaled and we can see they're really not going to be something that's worth having a million of. So even if Unry could do a million, I'd be surprised there'd be a million people that would want to have it because they'd realize it's not going to do a whole lot.

[01:17:11]>> It's when you get more and more capabilities that you'll start to see it. So we can see a scaling of maybe a factor of 10 per year from here out.

[01:17:21]That would be like the most aggressive.

[01:17:22]is probably going to be a little bit less than that. But the big thing comes down to the AI training is is that for some reason someone gets this incredible breakthrough that become very performative, people will be very excited. But at the same time, if the actuators are overheating every 20 minutes and the payload isn't very good and it's not safe and stuff like that, you get some sort of of of limits on that. And the big thing is going to be come down to the use cases. If you can find some really good use cases, even though it's got limited capability and it's able to perform really well in there, then you would see the need the um the reason to to want to scale what could be kind of a limited platform. I mean, >> the Model A and the Model T were not great cars by today's standards.

[01:18:06]>> Yeah.

[01:18:07]>> But back in those days, they were pretty darn good compared to the alternative.

[01:18:11]So, so that's kind of what we're talking about is that, you know, it it's it's when have we finally got to that model A model T level that it's it's useful enough versus what were really just like experimental kind of prototypes that people were racing around and people would have said that you know were too expensive, too difficult to maintain and and you know, you needed to have a mechanic and everything else. Not that the Model T, the Model A didn't need it either, but you know, they kind of reached a level that um you could start to see that scaling made sense.

[01:18:42]But then again, as you said, um you'll never know until you build enough to deploy. Yeah.

[01:18:47]>> For whatever cost you have in mind.

[01:18:49]Yeah.

[01:18:50]>> You won't find out what what wears out, what what breaks, and and they know everything is going to And I think this has been the maybe the surprise for everyone is they built them assuming they were going to be way more robust than they were and then finding a lot of components were failing when they didn't expect them to fail or they were failing a lot sooner and then it was like, "Oh my goodness, we're having thermal problems." you know, everyone was like that no one was expecting the thermal problems right out of the gate.

[01:19:14]So then is it wise to then design, field test and deploy a production line for a V3 at this stage for Optimus at Fremont >> depending you know this is this is the thing I kind of wonder is that um how much scale they need to have and where it is and it can always be that you know you are planning a couple things together you might say you know it's going to take us a while which is what Elon has said it was going to take them a while to scale and they may be saying that it's going to take us maybe two to three years to really get this scaled and that time frame we'll come up with a prototype because there's going to be a lot of components that are still going to be similar.

[01:19:57]>> Sure.

[01:19:58]>> And as they're doing it they're going to find out this is what we need to do to make those a bit more robust. It's not like they're going to completely scrap it. So there's a good chance there's like maybe there's 80% overlap between the current prototype and what you would have as far from the component level.

[01:20:12]And then you're going to start learning a little bit about the production. So a lot of times when when you start looking at a lot of the Gant charts you're putting on up there, you you want to be getting things that they kind of come online at the same time and not say, well, I'm going to completely wait for this before I start something else. Even though they're not really serial, they can be done in parallel. So if you want to compress your timelines, then you may be looking at that way. So, without them really knowing what their full-blown production schedule is and where they are in it, maybe it's premature. It feels premature to me to to be talking about a million line um already. But if they're really looking at like, well, we're going to produce a line that can only do 10,000 right now. Get that going, but we have ample space and now all I have to do is multiply that by what? 10 times then another 10 times and you can get to a million perhaps that's what they're looking at >> but then again as L said this is a highly modular line so they can just you know make changes almost on the fly I don't know if that's the right term to use but at the >> yeah I don't think there's really going to be a um a line that produces a million because that that's that's just crazy be coming off the line uh every two seconds it's going to that you you have sort of like they talk about these pods that are able to produce a certain volume of them per year and that when you say okay we want to double it and we're running it as fast as you know the way you double production is you try to run your lines twice as fast but at some point you just can't you run up to the limit of that one and say okay let's take it and clone it and clone it and clone it and the thing is um an automotive production line is just large because of like this the size of the product and everything else has to so on >> for humanoids and we've already seen if you look at u both you know the figure bought Q if you look at at at the 1x and those are also um been set up to have run rates of around 10,000 a year and that's not a whole lot of space you know so you don't really don't need that much space where a lot of your space may be actually is the component suppliers that are are shipping all the stuff and where you're doing the final assembly and so we don't know how much of that they're planning to actually do directly uh say in Fremont or in other facilities around there or the suppliers that would be bringing them in. But you know, assuming they're able to get ample supplies coming in in there, then it's just going to be a matter of taking that space and then cloning it a 100 times and then you get your million output, >> right?

[01:22:49]It's always fascinating to pick your brain. uh Scott and Marad and Gustav uh whose video is off right now, but uh we'd love to wrap it up for this chat.

[01:22:59]Thank you so much, gentlemen. Again, thanks to Andrea for joining us and uh we're looking forward to a lot more of these. So, thank you.

[01:23:06]>> You're welcome. And keep building.

[01:23:07]>> Thank you.