BONUS: Humanoid Robots Need More Than Servo Motors: Here's What

21:42Nikita Rudin:That is so cool. By the way, I believe, once again, it was, from what I saw, it was in a, the demo was not in the large, on the large screen. It was on the side. No, it was, it was on the large screen. It was. Okay. So I have, maybe I have a different view. Yeah. Riverside, man. I love this. We can, we can make robots walk and talk, but getting, you know, four cameras to line up, right? Yeah. I think we're all seeing different views is what was going on there too. And we just switched to this as our live platform like last week. And we're still finding our way around a little and stumbling. I appreciate you running through that again, though, so we can make sure people get to see.

22:26That was so cool, man. Thanks. What an interesting thing to do every day. Oh, you're good.

22:37Nikita Rudin:Oh, you got to switch the sound back. Well, we got a couple of questions in the chat, which is really... Okay, I think I'm back. Okay, you sound good. Great. Yeah, once again, getting that echo. Now we're all set. Yeah. So we got a couple of questions in the chat, but Corey, your immediate reactions after seeing that? My thoughts? Yeah. That was so awesome. It's the first time I've seen one that wasn't in like, you know, on a stage in some place that I'm not. I mean, this is absolutely, I'm not in Zurich, but I am very impressed. And I've seen the Unitree robots, big fan of Rizbot.

23:24And I'm just amazed at what's possible now compared to what was just a number of years ago, specifically in terms of like giving it a brain and teaching it to do tasks on its own. That's so amazing.

23:41Nikita Rudin:And with a little bit of imagination, you can kind of guess where we're going in the next few weeks and months. Yeah. Now the robot can pick up an object and walk across the office, can walk through doors. What we haven't shown, but it can also manipulate the box and can take things in and out of that box. Oh, wow. With a bit of, if you extrapolate a little bit, it's fairly easy to imagine how we're going to go to actual deployments in a warehouse or in a factory, things like that. And so the benefit of your system is that you can train it once with one of these robots, and then any robot who uses your system can then do the same thing, right?

24:20Nikita Rudin:Okay, so the whole idea is that it should be very easy to switch between robots and between tasks. Typically, it still requires some retraining. The key metric is not, if you want to train something or not, is how much human effort is involved in bringing a new robot to a new task. Because we're training everything in simulation and we have pretty good models of all of these different robots, it's very easy to switch one robot for another. If one robot can pick up the task, we can create another simulation with another robot and train it in the exact same way. If you wait a few hours and you have a new robot with, let's say, a new brain, a new specialized brain that can do it as well.

24:55Now, does that stand true as far as models go, or is that at the individual robot level? For example, one of the Unitree models. Once you've trained one of those, will it slap onto any other robots identical to that? So you might need the Unitree G1 version of the flexion tool.

25:23Nikita Rudin:Yeah, that's it. So we need to retrain it for different morphologies. If you change the motors, well, it needs to know that now my knees are much stronger than before. So I can walk in a slightly different way. I can now throw a door. Exactly. But then all the robots that have similar morphology, you can still change a few things. They can add a bit of weight. Our controllers adapt quite well. But if the robot is a different model with a different size, etc., then you need to retrain. Not the whole system, but just a small part of that system. okay we were talking before about an agent orchestrating the whole thing you only don't need to retrain that right the that agent just tells our controllers to go through the door that part is exactly the same but the actual motion of grabbing the handle and pushing on it will be different if the robot has a different size okay am i correct and i i'm going to say something here that might sound dumb and if it does pretend it's not for me uh um it feels like you're past some of the biggest challenges in this like like you're now at a problem of scaling these simulation rooms to teach new tasks and is that a correct assessment or is that off?

26:47Nikita Rudin:I want to say yes. I think honestly there are still many challenges ahead. But the overall system is pretty close to what it will be in a few years. That's awesome. We just need to expand the set of tests. For now we're doing fairly simple manipulations, grabbing boxes, maybe opening them. We're in Switzerland, so if you want a robot that can build a Swiss watch, that will still take a bit more time. Every single component needs to get better. What about a robot that can do alpine skiing? It will probably happen before the Swiss watch. That's actually really interesting. Is there any room for a robot Olympics?

27:30Because I've been watching the Olympics in Milan the last couple of weeks. So what about a robot version?

27:36Nikita Rudin:Actually, it is already a thing in China. Yeah, that's right. Oh, yeah, they did one last year. Happened a few months ago, yeah. Yeah, yeah. And it was wonderful and hilarious all at the same time. It was... We got some questions from the chat I don't want to ignore. Yeah, the question was, can you comment on the AI ability to learn from cause and effect so that it can cope with new environments?

28:02Nikita Rudin:Cause and effect is a very loaded term. Okay. Okay. It can be a very deep conversation. But I'll focus on the second part. How does it cope with new environments? Specifically, if we train it in simulation. It is a challenge. Overall, it's a challenge for robotics. How to make these systems adaptable enough to cope with new environments. Maybe I'll take a quick step back and start with the other way. If you have real-world data, how do you do that? The answer is pretty much you have to go and collect data in a ton of different environments. And that's how you make it generalized across different environments because you just collect data of everything.

28:42Nikita Rudin:So what does that collecting data look like for a simulation? Is that just like creating a bunch of different versions of the simulation or is it actually filming environments for new simulations? That's a great question. Without simulation, you literally have to go and build different environments or I think other companies are renting Airbnbs across San Francisco to have different lighting conditions. et cetera. In simulation, you're in control of the simulator, so you have to create different conditions and random. We're always talking about domain randomization, which means you have one simulator and thousands of robots training in it, and each robot has a slightly different set of conditions.

29:22Nikita Rudin:If we're manipulating boxes, well, obviously, you change the size of the box, you change the background, you add a bunch of objects, you create. You might need to create something that looks kind of like a room. You don't need to go all in, because another thing we can do is only show the right thing to the robots. I always use this example. If you want to train a robot that opens doors, and the naive thing is to try to create photorealistic images and simulations such that it transfers to the real world. But then you would need to create all sorts of doors, like any material, any handle, everything that exists, in all lighting conditions, and that's really painful.

30:04Nikita Rudin:but one thing you could do is use some of the other computer vision models to segment what you're seeing in real life for example there are many models that can segment the doors and then you for example you show it as as red to the policy and then the handle you segment it again and the handle becomes blue then suddenly all doors kind of look the same they're all red they all have blue handles so that's the only thing you need to be able to to solve oh that's complex interesting yeah so it's not exactly what we're doing but it's one way to describe how to bridge this seem to real gap by yeah basically you have to create a common representation that you can simulate fairly well and recreate in real life as well so do you see a future where the idea is i hesitate to say this not dissimilar to how a Roomba learns a room like at the time that you would bring this into your company or into your home, which is still a little ways out.

31:01I say that. Maybe it's not that far out. But is that an approach that would be the ideal scenario for it to be able to take a walk and look around and kind of learn where the walls are, where the doors are, where everything's sitting? I assume kind of like you mentioned how it did that in the room you're in, right?

31:23Nikita Rudin:Yeah, absolutely. That's how it would work. So you send the robot, either the robot walks around, or if it's easier, you can just walk around with any iPhone or any camera. Create some version of that map. So that map doesn't need to be super precise. You just need to have a general idea of if we tell the robot, go to the kitchen, well, it doesn't make sense for it to every single time relearn or walk randomly until it finds a kitchen. If you know where the kitchen is, well, just go to the kitchen. Some kind of a virtual blueprint of your facility would probably be wonderful to have inside it.

31:54Nikita Rudin:Yeah, exactly. And then the robot itself, we already know how to interact with a lot of different objects. We won't need to train it, or we are not training it to pick up every single object you could find in a kitchen. You just need to make it general enough that it can pick them up. Then once again, that's where the agents come in. Another example that I like to use is if you want to train a robot to cook a specific recipe, you don't want to collect data of every single recipe that exists in the world. That doesn't make sense. However, what you can do is give that recipe to Gemini or GPT-5, whichever number, some element, and break it down into specific subtasks.

32:36Nikita Rudin:And then each subtask will be, well, grab the knife and cut that thing, or pour that into the pan, or stir the pan, or something like that. Which are all separate skills that will be trained and the robot can execute them. Okay, I have a question. Oh, sorry. Go ahead. Finish your talk. I was just going to say that the whole idea is to combine the general understanding of the world from VLMs with specific physical skills that the robot can do. So which part of that equation, whether it's you, Affleckxion, or the LLMs, is going to be actually handling the memory, whatever memory that a robot has to have of a space?

33:15So it comes in and learns the space. How does it keep that latent and how does it keep that in its context? Is that from the LLM side or what you're doing?

33:24Nikita Rudin:That's a super interesting question. And there are many levels of memory. So even when the robot picks up the box, or let's use the example of the door, at the last second when it's pushing on the handle, it doesn't see the handle, because the camera is kind of blocked by the door. Right. It needs to remember that it saw the handle at that location. So there you need a few seconds of memory, and that's handled by our controller. But again, just even to know that you have to cross the door to reach whatever is behind, that's a whole other level of memory and that's maybe based on a map or a much longer term memory that would come from an agent.

33:55Okay. Do you think that's something that would be happening on the LLM level then probably or perhaps another system?

34:03Nikita Rudin:There are kind of two ways to approach it now. The most pragmatic way to do it now is to give some specific tools to those LLMs and the LLM can query. For example, we have a 3D printer behind the doors. If the task is to go to the 3D printer, the LM can query the map and ask the map, where did you see a 3D printer? And the map will respond. The more end-to-end way you have to do that is to have a huge context in the LM. So you feed the whole video of the robot walking around and then the LM has to figure it out on its own. We're not quite there yet in terms of efficiently using that context. Memory will do a lot for that, I imagine, over time as far as where that could go.

34:43We have a really good question from the chat as well. how long does the battery last and how and i realized that this is not your robot but i would say on both for that matter and is is there a future where it can hot swap like a dewalt drill

35:04Nikita Rudin:or something oh absolutely so the the battery on that specific robot can last i think about an hour and a half okay yeah unless it's doing backflips and i think it goes it goes down a bit faster So walking around, around an hour and a half. Other robots have larger batteries. Something like four to five hours is achievable. And then hot swapping is definitely a great deal. That specific robot cannot do it. Others can. Quite a lot of them have two batteries. And you can hot swap and you remove one, replace it, and then remove the other. I was just thinking that made a lot of sense to be able to have it be like, I'm about dead.

35:41I guess dead feels sad right now. A nice robot. I'm tired. I'm tired. Yeah, I'm sleepy. I better go get a new battery. And yeah, and I also want that feature for myself. Oh, my gosh. What about the, so there's another question. How does the robot know what to do? Are you giving it commands? You kind of talked about that. So what you were doing during the demo is you were giving it essentially like typed commands or a button, like a preprogram button, which the agent could then do. like in the situation where the LLM is hooked up to it, right?

36:18Nikita Rudin:Yeah, that's exactly it. Either we specifically trigger a skill if we want to show it live on camera, or we go through the agent and the agent understands the context and trigger and does the same thing. Yeah. Would you imagine then a robot in production having like, call it a menu of skills or like a library of skills that it would be calling and using? Or would it be more on the fly kind of deciding? That's a great question. It depends a little bit on the actual application. In general, whenever you have a specific menu of skills and you can only switch between them and you cannot interpolate at all, you might find situations where it's limiting.

36:56However, having said that, if you're on a production line and you're doing the same

37:00Nikita Rudin:motion the whole day, you don't necessarily need a crazy adaptable system. Now what's interesting is even if the robot is doing one single motion and you can remove the whole agent thing, all that, training that motion is still easier if you already have a platform that can do a lot of different things because you're kind of just fine-tuning it for that specific motion. Whereas if you're designing a whole system just to solve that specific motion, it will be much more inefficient. It will take way longer to achieve the same level of performance. Interesting. Why is that? That's so interesting.

37:33Yeah. Yeah.

37:35Nikita Rudin:Anyway, because you're starting from scratch. It's like, you know, now we have so many LLM wrappers applied for different use cases. You know, let's say if you want to solve translation, now it's very easy because you already have an LLM that is very good at it. And if you start from scratch, well, you have to design a whole new system for translation. And you probably wouldn't come up with an LLM that can code as well, because it doesn't make sense to do that. Yeah. the exact same thing applies to robotics okay it's easier to start with it from a general system and then make it fine-tune it for a specific use case yeah that makes sense then to start from zero for a specific use case okay that makes sense i know nvidia and and i'm not if if i tap on anything here that's a little proprietary or something please feel free to to say so, has released a pretty large mountain of robotics data sets and models and things.

38:33And it seems like I continue seeing more and more robotics focused models, research around robotics and AI. And it makes me wonder, like, how do you decide what are we going to use for this? You don't have to say what you use, but I feel like trying to figure out where the starting point is. Are we fine-tuning a model? Are we building a model? I feel like there are a lot of decisions to make there.

39:05Nikita Rudin:You're right, and it's actually pretty hard to keep up with every hardware or software development happening in robotics. I think every single day, my X or LinkedIn feed are full of different demos from all over the world of robots doing cool things. Yeah. I think the key is to be adaptable enough to be able to quickly incorporate. When someone shows that a new tool is very useful for a specific thing, we want to be able to incorporate the specific tricks and techniques that they use. Okay. It's very hard. I'm not in favor of saying our bet is that that specific thing will solve everything because then you kind of ignore what everyone else is doing.

39:47Nikita Rudin:You have to have a bet. You have to push for a certain direction, but you also need to be flexible enough to quickly adapt once someone shows that something else works as well. I think that's even a good approach to just AI usage in general for people. Grant and I have talked about it before that, you know, some tools are better at certain things and that changes every day. Like, you know if if what you need is the best of the best right this instant you need to be prepared to move around a little you know yeah that's a great that's a great metaphor that's right um do you have a particular model or type of model that you use on the llm side or that you think is the best for working with robots or is it just whenever the best new thing comes out you switch to that and test it or are you testing all of them so it's an interesting question so we we switch so Also, there are two ways we use LLMs.

40:38Nikita Rudin:One is really for this high-level agent. We're usually talking about three layers. The first one is the agent, the one we call the command layer. It goes pretty much from text to text. It's really what VLMs are really good at. And the second part is we reuse pieces of an LLM and create what we call a VLA, so it's a Vision Language Action Model. And that one outputs specific commands to the robots. It's not going to text anymore, and it really has to be trained on robot data. Both of them reuse pre-trained models. The VLA gets used. The vision encoder is pre-trained. The language encoder is pre-trained.

41:17The action part today also can be pre-trained

41:20Nikita Rudin:for models released by NVIDIA or others, or you can also train them from scratch. The VLM part, we start from open source models, and then we fine-tune them to make them better at understanding what robots can or cannot do. That's cool. What's your compute bill look like? I'm so sorry. Oh, it's got to be massive. Honestly, it's surprisingly low. Oh, that's good. Mostly because we're training things in simulation, which is, again, surprisingly effective. So we can simulate thousands of robots on one single GPU. Just with just a few GPUs, we can actually do a lot. Is there any sort of scaling law with what you're doing with simulation, where if you expand that to 100 GPUs, you'll be able to do this way faster or you'll be able to unlock something different?

42:10Or does it not work that way?

42:14Nikita Rudin:Interestingly enough, there's something I explored in the very first paper of my PhD, which was a very long time ago. I didn't call it scaling laws back then. I'm not sure it was a thing. I should have. Certainly not a common phrase. You could have coined it, yeah.

42:32Nikita Rudin:And the reality was that it stopped scaling at some point. So just throwing more data at these reinforcement learning algorithms doesn't make them train faster. The problem is that there is an aspect which is exploration. So you need to try something, get some data, and then update your model, and then try something new again. And if you just throw more data without having this update step, it doesn't help because they're just more of the same. You really need to have this iterative loop. So we call an on policy algorithm, which means that the data has to come from the current state of the model.

43:06Nikita Rudin:We cannot reuse huge data sets from before. Hmm. That's really fascinating. Is that, and that's unique to like vision language action models or robots in general? That's very specific to reinforcement learnings, not necessarily to VLA's, but to reinforcement learning. It's one way to the reinforcement learning. there are also other ways this one seems to be the most the most efficient at least for robotics when doing simulations i'm curious is the actual figuring out of how to do it in a simulation phase quick and then there's a ton of of like kind of committing it to memory and getting good at it takes on what what end is more difficult is it is it find it figuring out the task or is it getting consistently good at it one challenge specifically with reinforcement learning and it's not specific to simulation but i guess it's a reinforcement learning thing um something's really hard to find a way to to solve the task but sometimes that's not really the way we we meant again there are many ways to for a robot to open a door it can push the hand it doesn't with its any part of any part of its body you don't want to carry a coffee mug with one finger through the hole for example yeah i'm just imagining the goofiest scenario it's like like opening the door with the foot or yeah like like using its head to do something it's like no i mean yes that's effective but that's not what we want it is an option it's just not the proper option yeah Yeah, exactly.

44:47Nikita Rudin:And so a key thing is how to steer it towards the right solution. That's actually where a few human demonstrations can come in. So if you have a video of a human going through a door, you can reuse that to not to tell the robot exactly how to do it, but just to steer it in the right direction. Don't use your feet. Your hands are good at opening doors. That's so amazing. I never thought of that. I remember when I was writing a newsletter that was all about like the business news will impact the future five years from now, like back in 2022, there was a paper that came out of ETH that was all about how like, hey, you could just like show robots videos of people doing things and then it can learn from those.

45:32Is that still true in terms of the simulation that you're doing or the simulation more like the data that you produce from simulation more in depth than just videos?

45:44Nikita Rudin:It's definitely more than that. So one thing that's missing in videos is you won't get to send commands to the motors. You have no idea what to do with the motors of a robot to actually achieve the motions that a human does. This is the easiest thing to understand. It goes a bit deeper because the robot doesn't move in the exact same way a human can. If the robot could replicate all the motions of a human, that would be much simpler. We have to do a lot of tricks and be smart about how we actually use the motions of humans to then there's a concept called retargeting that's one way to do it to you take the motion of a human and you retarget it to the motion of a robot so it would be almost like translating it into what the robot could do is that kind of that's so they're kind of let's say two steps first you translate the motion itself to what the robot could do talking about kinematic motion so you have no concept of forces in the motors or etc and then you actually train a reinforcement learning policy to execute that specific motion and it learns how to control the motors to do it.

46:46Nikita Rudin:So there are still a few steps you need to take to use human videos. Okay. Another great way to use human videos is not going into the control of the robot, but again towards the agent. So like just training the agent to be better at inferring, I guess? Exactly. To understand how to break down a task into clear sub-tasks. And then, there, if you do it right, it's much easier to translate between humans and robots. Because, again, I'll put open the door. It doesn't matter if it's a human or robot. As long as both understand the same language, it translates from one to another. Okay. Here's a question I'm just curious of.

47:30Is there a task that was harder than you thought it should have been? and one that was easier than you expected it to be.

47:44Nikita Rudin:I won't answer exactly your question, but I think it's still an interesting thought. It seemed like the motor control, bending over, grabbing things from the ground or walking on stairs and things like that should be easy because they're fairly easy for humans. And doing long-term planning and understanding that you have to go through doors and you have all these different sequences of things that seemed or i would have expected that to be hard but the exact opposite is true our current vlms have pretty much solved the the agent and the reasoning part not exactly we still need to fine-tune them to really make them work reliably but it's really really close however but the groundwork is there already yeah exactly okay okay and you can really change the prompt and the exact same the robot does something else in the new environment.

48:34Nikita Rudin:So the generalization part is kind of there for that. The motor control is still challenging. That's right. Okay, we got a really specific, interesting question. You can take it any direction you like. What about a virus in AI models? How would that play out in a robot scenario, and how would you, I guess, protect against that? I don't know if that's a real scenario, but but this person really wants to know. It could be. Security is a very big deal once you have many robots out there. It's something that has to be taken very seriously. For the honest answer is we're not exactly there yet, this is the bottleneck that we need to solve right now, but it can very quickly become one.

49:20Yeah, probably more so once it's in production and threat actors know like, hey, a big surface to attack is like this, you know, BMW plant where they have, you know, hundreds of thousands of robots working or whatever.

49:32Nikita Rudin:Yeah. And there's a more, an underlying question there is, are these robots connected to the internet or not? What's your take? That's a big one. It's a big challenge. Everything would be so much easier if they were connected to the internet, because then we can use much larger models, run somewhere in the cloud, we're not limited in compute, etc. but typically in industry they cannot for security reasons. So we need to find ways to still run large enough models that cannot fit on the robot itself. For example, somewhere in the facility, but not in the cloud. So what's the bottleneck there for edge computing, do you think?

50:13Or I guess the stronger bottleneck, because these could both be bottlenecks. Is it the hardware that can run on device, or is it the model that can run on device?

50:24Nikita Rudin:Well, I guess both, because the hardware that runs on some device is not strong enough to run the models that we would like to run on the device. And that's another reason why we're split... Sorry. Yeah, they both have to work together, and that's the reason why we are splitting our whole system into at least three layers, because it's three neural networks of different sizes that also run at different frequencies. The part that controls the motors has to run very quickly and really onboard on the robot, because the latency will just make the robot fall. Same for the VLA, it has to run on board, but it can be slower, so the computer can take a bit more time.

51:00Nikita Rudin:And then the agent part typically is really too big to run on the robot, but it's also not a problem because we only need a response once every few seconds, for example. So it's fine if you have to send a request and get the response back. But I guess technically it wouldn't be fine if you couldn't connect it to the internet. Well, you need a server rack somewhere in your factory that can run those models. Oh, okay. So you could do it over like a LAN or something. You can have Wi-Fi internally in the facility, but it's not connected to the outside world. Got it. Is it specifically beneficial that it would be the internet at large, or is it possible to have some kind of a, I don't know, a gateway to a private network or something or a, you know, kind of a, you know, maybe it's child safe controls, as silly as that sounds.

51:55Maybe it's the answer of like, here, you get the youth browser with all of the safety protections and the other things that... Instagram teen accounts. Yeah, you get an Instagram teen account, yeah.

52:11Nikita Rudin:I'll try to sell that to our clients. Don't worry about security. We're using a child safe. That's right. They can watch Netflix. But that's kind of like what you were describing, right? Is that you'd have like an on-prem sort of like, you know, not internet, but a personal net essentially. Yeah. I think all of these are options either onboard computers gets so good that it's really onboard the robot. I think that will take a long time or it's on-prem or it's maybe the building next door, but it's a larger compute center. So all of it is possible. Before we let you go, I have a question I have to ask.

52:50And it's just sheer curiosity because I enjoy asking people this. What's your robots and homes, robots broadly available in businesses timeline? line if you just had to guess based on what you know, what you see, the companies that are working on things?

53:12Nikita Rudin:It's a very common debate right now is are we going to get robots in homes or in industry first? People disagree. I'm in the industry first camp. And also I think industry matches better, is a good match for what we're, the kind of technology we're developing with simulation where we can still create specific simulation for a specific production line. it's not completely zero shot it takes a little bit of effort but not years it's maybe a few days for an engineer yeah and then we can deploy 100 robots doing that that's it it's worth doing it well in a home you cannot send an engineer to every single home you're going to send a robot to yeah so i think industry will happen first and people like to ask about when is the chat gpt moment of of robotics i think unfortunately it won't be one single moment the i think the most obvious reason why is because you need to build those robots and will take time it's not software then you can just copy paste across everything.

54:04Nikita Rudin:But even software will take a little bit more time to adapt to all those different tasks, different factories, different warehouses, etc. I think we'll start seeing robots deployed between the end of this year and beginning of next year. Okay. You'll have to go look for them. It's not like you'll see them everywhere around you. Yeah. Then it will progressively expand. It will expand very quickly. We see these little... I don't know that they're actually robots they're definitely not humanoid robots but these things in stores that are like now my grocery store has this six foot tall skinny bar on wheels that walks around and it'll talk to you but it's scanning and doing inventory throughout the store as it goes through it'll say excuse me stuff like that and our gas station quick trip down the road has this one that is constantly walking around cleaning the floor and saying excuse me I'm cleaning the floor and he's got eyes that blink.

55:00And, you know, and I think those are the first, while that is very different from this, those are the first, like, out in my normal path of life going places, I've seen these things. And I always think, I guess this is the very, very rudimentary beginning of what's to come.

55:26Nikita Rudin:I think that's right. So next step, you'll see the same robot in different places doing different tasks. Like, oh, actually, I've seen that exact robot doing something else in the store or the gas station. And then it will expand. Then it will be completely normal to have multiple robots walking around doing all sorts of things. It will be strange once you'll see a human doing a very manual and labor-intensive task. Yeah. Oh, man, there's so many branching paths from there I could think of. But one reaction to what you said, it's almost like the chat to PT moment for robots. Maybe it won't be a chat to PT moment.

56:00Maybe it'll be more like the iPhone, where Steve Jobs came out and showed the iPhone and like, oh, wow, this is really cool. But you didn't even realize what all that meant, like with the app store and all this other stuff until, you know, a couple of years later and you see how actually revolutionary it is. Perhaps robots go the same way.

56:18Nikita Rudin:And it's I think that's a great example. And not just the iPhone, but the whole Android ecosystem as well. A few years after the iPhone revealing, you have all these different phones, all these different devices that can weigh more than just the iPhone itself. Yeah, we're kind of still at the point where we're seeing them in demos, at special shows, but there's yet to be the person who walked on stage and says, it's right here, it's$8 ,000. You can order it at apple.com. Go get one. You know, we haven't hit that thing yet. And I think that's really exciting and curious. And I tend to agree that you're right about industry, too, because, you know, the truth is early on, you know, the earlier you buy, the more you're going to pay, honestly, for probably less technology, as is the way that most tech works.

57:10is that if you bought a laptop in the 1980s, you paid five grand for one gig of RAM. I mean, excuse me, one megabyte of RAM. It's going to be that way with the hard drives. Megabytes of hard drive, you know. With the way memory is going right now. I have one more question, but then we're going to drop here right at the hour mark, so I want to make sure we sign off and you can let people know where to go to check out Flexion. um are you i guess are you bullish on the idea of a ai like post abundance like like or like an abundance utopia with robots like where we're not working or the robots are doing most of our work or do you think that that's perhaps far far far away

57:54Nikita Rudin:I think it is pretty far away. I cannot give you a number of years because again it will be progressive. 50 years from now probably we're there. It's going to be much shorter than 50. I don't know if what we'll have in 10 years will qualify as a Robototopia or not. As long as Earth doesn't look like Tatooine by then I think we'll be okay. Well, at least Zurich won't look like Tatooine. Zurich's too pretty it's too pretty well Nikita where can people go to learn more about what you're doing at Flexion we're not super active but we're publishing updates on X and LinkedIn we have our website flexion.ai that's the best way to follow what we're doing go watch their demos absolutely watch their demos there are like two right now and they're both really really good and I had good intentions of showing them while we were on here and I completely forgot earlier.

58:55But we'll absolutely show them. We showed a live demo, so that's better. That's better, yeah. We got a live demo, which is way cooler. And there'll be more coming soon. Excellent.

59:05Nikita Rudin:More demos coming soon. Nikita, thank you so very much. And hopefully we'll do this again sometime down the road and have some other cool stuff to see. Thank you so much. This was really exciting. I'm glad we got to do the demo live. Excellent. Me too, me too. We didn't know if it would happen. Well, have a great day. We really appreciate you coming, and we'll be following. Thank you. Bye. Bye, everyone. All right. That was so cool. Rant. I believe we're still alive. That's okay. I don't care. Actually, I'm glad we are, because now I want to talk about it with you, because that was neat. There were a couple of questions that I wanted to hit at the end, but I wanted to respect his time.

59:52But one of the things I was wondering is how, if at all, the current memory crisis will impact robotics. You know how what's going on right now is essentially that the price of RAM and HDDs and now CPUs, all of that is going, basically the supply of all those things is drying up, and therefore the prices will increase. I wonder how that's going to impact robotic development. also I wanted to ask more about how their backpack system works which I think is really cool we're going to have robots running on 80 ,000 1 megabyte RAM chips left over from 1993 that's what we're going to do look I mean in theory if you have an efficient enough algorithm it should be able to run it that's what I need I need a rig in the house with all of the 1 megabyte RAM chips that exist we'll buy a gajillion of them if you buy Richard Sutton's argument with the better lesson, which is that an efficient algorithm will just scale perfectly with compute, then, you know, like, we don't have the current LLMs are not that.

1:00:55You know, so perhaps that was something that could run on that type of system. You know, a thing that was interesting about the virus question for me and that I was going to call out and forgot is that, you know, you watch these videos and they look like these friendly little robots running around, but I mean, like, it's important to know that these things are made of steel if you've ever watched a clip like on instagram where one has a glitch and starts you know spinning around and kicking and flailing at everything like you don't want to be anywhere near that like it's gonna break bones yeah no no those things that that that that is very scary which is why the work that what nikita is doing is really really important um to be able to train these things so that they're safe and in environments exactly yeah that was an interesting answer too about like oh yeah you could just have like a local internet um that it runs on that's like safer and i was like yeah that makes sense i forget you could do that yeah that's the thing i guess uh yeah yeah i thought i thought the whole the whole kids internet thing was you know i mean you could you can you can you can gate it off a bit i mean you know the truth is we're almost at the point where there's a will there's a way though the fact of matter is you know and there will always be bad actors like that's not a thing that's gonna go away anywhere in the near future i wouldn't think well i was gonna say i think the the kid solution actually makes more sense in the near term for like agents like we should yeah we have all these agents like an agent specific like like that's what i need to do like account type like like we they need to be like okay don't take this the wrong way but but like second class like citizens on the internet or something like that where basically like we have them like if they're going to be on the internet they need to identify themselves and they need to have their own sort of like protocols and like yeah you know i would love that because i wish an awful lot of tokens downloading bloated HTML and ad graphics to run up as tokens.

1:02:56I would much rather have a really simple approach to that. Well, WebMCP is a Google solution to that, right? Like what they're doing with WebMCP. Do you know what happened? Remember me a week or so ago talking about the situation where my wife starts texting me and wants to know why the computer's talking? And I go down and I couldn't find the browser. Well, I get off work day before yesterday and I go downstairs to my desktop where my open claw little multi-buddy lives. And I open it and it's running the Brave browser actively. And it has more than 50 tabs open. Half of them read it. it's learning the other half of them a half of what's left was youtube for for people who don't know open claw is the agent that you basically give all of your permission and connections to and you just let it like live on your computer and it has a heartbeat and it can act on your behalf so what cory is saying is that this open claw agent was just opening reddit tabs and reading them yeah just for funsies i swear to you i came down and it was open and and the bar was so tightly packed with tabs across my 19 inch monitor down there, which is my little monitor for my easy chair and that you couldn't read a letter on any of them.

1:04:29It was just corners of tabs all the way across that screen. And I was like me when I'm looking at the news for the day. Yeah. And then I go and I look at the tokens for the day and I realize it's mowed through. I don't know, some obscene number of, you know, 20 million or so something like that and it's just like oh my god so yeah i did a lot of neat stuff to fix that last night but but i don't want to hijack the robot talk so yeah yeah that's fine unless somebody asks if someone's interested i'll get into what what i did yeah yeah well we can yeah we can definitely get into that well yeah we have a gab about whatever do you have any other reactions to the chat the other thing i wanted to ask him that i didn't have that we didn't have time for was you know he mentioned humanoids you know or you mentioned the scenario where we're seeing like the robot more robotic looking robots um you know walking around i was gonna ask him what his timeline is for when we'll see like kind of more human looking robots walking around where it's like almost like westworld where you almost can't tell the difference between a human robot like where you go to a ball game and there's a non-zero number of them and the guy just like looks at you or just like you know roaming around doing service work oh yeah that's more practical i was imagining you're in the more practical approach is that it's rolling hot dogs not that it's watching baseball no i was just imagining this is a robot sitting next to you and it's watching the game with you and you're like i don't know if this guy's real or not you bite a beer it's like dead internet theory but in real life the practical approach is that the second the second major league baseball can put one on a pitcher's mound they'll be down yeah well he mentioned the robot honestly robot baseball robot sports wouldn't be so sick will be so sick okay so first of all there is a robot olympics i forgot about that um that china did one uh second there's robot boxing in san francisco have you seen this speaking of which when we are in town in a few weeks, we need to find out about robot boxing, robot fight club, the things like that, because I would say that.

1:06:35I don't know. The robot fight club stuff I've seen, I know this is going to sound bizarre, left me feeling a little guilty. Well, let me show you some pretty wild stuff. Hold on. Now it's going to get real. Yeah. Okay. One second here. I'm going to come off camera so I can cough like a monster. Ika, I promised you that we would probably only be an hour, so if you want to drop off here, you're totally welcome to. Totally cool. We get it. Okay, so Robot Fight Club. Let me see if I can find footage of this. Robot Fight Club. Okay, they've got one in Austin, too.

1:07:21So these are two of the unit tree robots that we saw earlier.

1:07:32They don't think they can knock each other over. five four three two one one so so that one leaves a little bit to be they're not all comically bad either no no like on occasion they'll be rough uh and you know i don't know what it is because i don't have any guilt with battle bots in fact the only emotion i have with battle bots is that i want a battle bot but why do you feel guilty if it's two humanoid robots fighting versus two little you know what it is i look at their motion and it's a little like we're watching toddler fight club or something and humans are so excited about it there's something very dystopian about humans watching another i'm gonna say species because i don't have a better word for it uh kind of brutalize each other well it's it's like the classic uh thing where if there's four if you're on a street corner with you know looking across and there's four different corners and on one corner someone's playing soccer and on another someone's playing basketball and on another someone's playing baseball and on the fourth one there's a fight which one are you gonna watch yeah you're gonna watch your attention's gonna go to the corner with the fight yeah absolutely because nothing says human more than loving a good train wreck Yeah.

1:08:56You know, the other thing is, I think we need to acknowledge how much Unitry is winning in the robotics field right now from a consumer perspective. Well, yeah, they have an affordable robot that anyone can use. And another company released one this week for$79.99. And I have it in a chat. I'll find it. I'm looking in Slack. Give me just a second. Yeah. Where is the$8 ,000 robot? an actual fighting tournament oh this is cool i believe it came from our cyber security guru ken underhill but it might not have been no it was in daily tech insider one of our sister newsletters hang on just one minute uh that's where it was daily tech insider which if you're looking for just general tech news every day with a little bit of a neurani spend you should absolutely check out DTI.

1:09:55It's a fun newsletter. It's run by Justin Myers. Justin is a wonderful human. Just a genuinely good dude. He's a hard worker and he takes everything he works on. Very personal. Oh, oh. Oh, heck no. He's fighting two at once.

1:10:18Oh, it's so funny. Oh, you see how the ref had to get out of the way there because that would have hurt. Have you ever seen the one where people are like, there's like robots in the ring and there's a guy in there trying to ref him and he's working with one of the robots and this other one is walking around behind him like this and occasionally walks up and like kicks him in the thigh and he shoves it away? No. No, no, no. I missed that. Look up Weave Robotics. $8 ,000. This is the one that Justin shared out today. This guy? It's stationary, but it's a laundry folding monster. Let's see it. It's$79.99, probably$99.

1:11:03I know why you like this, because the face kind of looks like Johnny. It's got a Johnny 5 vibe, doesn't it? It's got Johnny 5 hands. I really, really love it. So I'm going to read what Justin has here. Weaves 8 ,000-later laundry robot. Still needs humans. Alexa can set a timer, but Isaac Zero might need to phone a friend. Weave Robotics has started shipping Isaac Zero, a$7 ,999 stationary robot that promises to fold your laundry while you bend your latest soap opera. Drop a load of shirts, hoodies, pants, or towels into the Bay Area-only machine, and 30 to 90 minutes later, you usually find tidy stacks waiting for you.

1:11:43Wow. Have you seen Physical Intelligence? maybe i know i know the name i was exposed to a bunch of these physical ai companies at aws because it was kind of a sizable focus they had a lot around agriculture and stuff like i i played with agent agent agent which was really cool a-i-gen a-i-g-e-n is the name of the company i always pronounce it wrong guy named kenny owns runs it developed it really really cool guy These guys are cool, but they have this weird little robot that's almost like a little crab dude. You see this? He's got these antennas. But he can do a lot. He can do quite a lot. I'll be honest.

1:12:26That's perfectly acceptable for the various tasks I would love. I mean, I don't guess I need someone peeling my oranges for me. I definitely have them make me a peanut butter sandwich. Can I flip a grilled cheese? Oh, wow. This is so cool. That's my other theory on robotics is if people are so scared of some Terminator possibility, is really the first thing we should do, drag them into our homes and make them do our tasks? Like that's not a thing that ages well for any level of society in the past when we've done it. To the point about security, right? Like what if you had a robot like this in your house and it wasn't secured on your local network and instead it was, you know, connected to the internet or worse, open claw and anyone could message you on Discord and like hack your robot?

1:13:19That's what I need. I need a robot that my open claw can direct. So no matter where I am, if we're away from home for the day, I can be like, hook up Penny, take her outside. I'm not home. And I want to make sure she gets to go or turn on the television for her. Give her some puppies to watch. I actually think so. So for people who don't know, alignment is one of those things where basically we're talking about how to get the robot to make sure the robot is like good. Like, like basically like that keeps humans. That's basically it's safe, like it operates safely. It doesn't operate where it wants to harm humans.

1:13:56You know, how do you make sure that it stays on policy, all those different things? I actually think alignment could potentially be an easier thing to solve for. the security issue I think is actually harder because you have to come up with new novel ways of securing devices that are talking to each other. Yeah. By the way, this is the Robot Olympics in Beijing from last year. So here's some boxing. I'd love it. Oh, wait, we're going to actually touch. Oh, got him. Yeah. Look at that there. Michelle Tyson. Yeah. I'm giving it a shout. I couldn't call it Michael in pink for some reason. I felt the need to gender the robot.

1:14:41That's the patriarchy for you. Oh, yeah, the running. If anyone watches the real Olympics right now, shout out in the chat. Oh, it's like this. Oh, no. I've been into it.

1:14:56Let's see if there's more videos like that on YouTube. They're pretty funny. I have. that is really cool by the way uh terabyte here is asking what's the best way to get in touch if everyone shows our ai uh for grant that is probably is it team at neuron daily dinner on daily.com yeah you can email us at team at dinner on daily.com and uh we'll check out what you got um and also uh if it's cool we can have it on a stream sometime yeah and uh you know also i'm I'm easy to find on usually Twitter or LinkedIn at Corey Knowles. C-O-R-E-Y-N-O-L-E-S. Hit me up anytime. As I'm usually out nerding about on something.

1:15:43There was actually a pretty big robot. This is the robot fast walk competition. Oh, no. It nailed the wall. It crumpled. If that was a human, people would cry. Yeah. You'd be like, oh. Well, you know what sport is pretty intense like this? I don't know exactly what it's called. It might be pursuit. But it's the one where there's like four guys and they're all ice skating like really close together. And like they, without fail, like they do like 15 heats or something. And like they knock into each other like four, like every fourth time or like four times in a row. It's really quite, it's like this.

1:16:20This is amazing though. It is so a little sad. but also hilarious. I see what you mean about the toddlers. Maybe this is how parents feel when they watch their kids' soccer game. As a guy with four children, I can absolutely tell you this is what kindergarten soccer feels like right here. Only the difference is that robot just got one and a goal. Yeah, I was going to say, I played soccer for four years and I never scored one goal so these robots are playing better than me. The robot just scored. AGI achieved, perhaps? What was the thing I was just saying? Oh, there was a good... Let me go to Reddit.

1:17:01Robotics. Oh, that's a fun deep dive. Yeah. This is cool. Have you seen this? So this is, I believe, part of a presentation for Chinese New Year. Yeah. This is freaking wild. Let's full screen this. Yeah. Are those unitries? Oh, yeah. I think they're G1s. They might be a different model. Okay, now the G1, I believe, is the$22 ,000 model. Yeah, this could be a newer one. I can't tell from the frame. But it's very much the G2 legs, the G2 face. That's what's really distinctive about them. They've got that concave face with the blue LED. And think about this. They're putting kids on stage with these robots.

1:17:48these robots have to be perfect for it to go on stage with a kid because if this accidentally fracture their skulls yeah like if this kick like this nunchuck actually like hit this kid like that would be so horrible so think about how precise they have to do to be able to do this but what if this kid hit the robot with the nunchucks i mean uh it depends on how strong he is you know i'm just kidding oh like look at this there's a guy with a sword hold on hold on oh now Now it's a party. Robots with blades. That's a good idea. Okay. I wonder if I can watch. Let's see if somebody posted. We should go look at their models here in a minute, Grant.

1:18:27Sure. But at what cost? Just kidding. This is amazing. Yeah. Well, yeah, you know, people on our robotics are fine with it. You show this to anyone on Blue Sky or any other platform, and they're going to be like, this is horrible. Yeah, different vibe. The Terminator is real. Oh, this is interesting. Okay, so here's the Spring Festival Gala performance from last year. Full minute version of this year, so we can compare them. Okay, so which one's this one? This one is from a year ago. This is what they could do a year ago. Okay.

1:19:12It's like fine. Yeah, it's still impressive. Yeah, this was impressive a year ago. I remember when we posted about this. Let's go watch Atlas. Atlas doesn't get enough love. I've got it pulled up, actually, in another tab. Do you? Yeah. I think we don't talk enough. Grant, let's get somebody from Boston Dynamics on here. Why haven't we had Boston Dynamics on yet? We should do that. Well, you know, we're just opening the floodgates for robots right now. Well, I mean, I think it's a greener. This is cool. Wait, did you see this? Look at this. They're like spinning the and catching them. Wow. That's cool.

1:19:51They look like, you know, those plate mat things. Like placemats, that's the word. I'm assuming they're not. That's a year ago. Now, oh, this is not available. Oh. Let me see if I can look it up. There was a fake video of Atlas this morning that I almost bought hook, line, and sinker dancing. Not me dancing, Atlas dancing. yeah imagine with c dance 2.0 you can make some pretty convincing looking pictures yeah and the funny thing is uh atlas is impressive enough i don't think i i don't think you need to get it to do things in there that is not uh great okay here we go i think this is it this is the one from this year

1:20:39whoa whoa whoa whoa okay hold on we're gonna we're gonna we're gonna full screen this so here's the kids right they're hanging out and then now we meet the robots here

1:20:58okay

1:21:03we've seen a little bit of this the the movements are so smooth it's crazy yeah i mean of course i can't stress enough this is the most affordable robot company actually i don't think it's the most affordable but it's one of it's like the most affordable best at an acceptable level of quality i would maybe is the if like what you want is a real doggone robot. This is a real robot. they have, I want to say it's a$6 ,000 or an$8 ,000 model. Maybe even a$5 ,000. And they've got one in the low to mid-20s. This is just so cool. It is. I wonder how many times in training one of those kids got hit by one of those.

1:21:54and like watch watch their feet okay yeah just just sit here for a minute and watch their feet watch how they touch they're identical they're so synced i mean there are still like some balancing things going on but this is just amazing it's it's amazing it is i mean obviously this is all pre-programmed right yeah so i don't care it's put it in a real fight and we saw what it looks like it looks like they're like boxing each other like missing but it's still so cool this i'm excited to watch so check this out so they're like jumping over stuff

1:22:33wow like what what that looks human they just ran backwards and flipped over it this would honestly be so fun if you're one of these kids if i were a guy who was offered a super bowl halftime show i would have these things on stage with me i would watch your super bowl i appreciate it i appreciate it i want to skip through because i want to see the guy with the sword here we go stand up and read matt schumer's paper this looks like a new model you see this one oh yeah it does that one in the center maybe that's a prototype perhaps man if you're off by an inch that kid's hand those kids hands are broken man yeah they have to be perfect to do this actually do their hands look different do they look a little softer oh they yeah i think there's perhaps a foam thing on there that makes sense yeah i don't know if it's different it's not metal i mean i mean i say that it's almost got to be at at least a little because, hey, thanks, Jeff.

1:23:43We appreciate it. We have so much fun. Appreciate you joining. By the way, for anyone who's here, if you haven't yet, please take a minute to subscribe to the channel. We'd love to have you around more often so you can keep up with our stuff. And, oh, there it is. This is the one that just came out. Boom. Speaking of Robot Olympics.

1:24:06That's the warrior right there. I see this thing and I think, can you imagine 20 ,000 of those coming into your city? Oh, gosh, no. Oh. Look, it's not perfect, okay? Even robots have off days. Oh, no. It's still closer to a somersault than I am. Oh, gosh. And then you see the parts fly off. That's funny. I like that they put this out. I appreciate this. Same, same. Yeah, because we only see it at its most prime and perfect quite often. Yeah. But obviously they stumble, they fall, they break. What an expensive thing to break, too, by the way. Yeah. And all these parts are custom, per my understanding.

1:24:59There's really not a supply chain for this type of stuff. Watch it land that. Boom. But then, because you land this one? Okay. Oh, OK. Fair enough. Actually, that's not nearly as good as the one before. Yeah. Before it like it like. I think this is like the progress, right? Like how long it takes to go from like falling down to the finished demo that you see. I love the round head with lights. Something about it is a little terrifying. Yeah, it's form factor is great. It's wait, wait, wait, wait. Whoa, whoa. Did that say that you could get the quad for four grand? this guy unitry has a quad for sixteen hundred dollars I mean if this is to be believed I don't need to know that I don't like these ones as much though because I find them not as useful I mean they're good for like surveying like if you had a construction site you could have this like surveying the site at night or things like that or I guess you could have it like there's a dinner tray on it and it could walk out and bring you your dinner I also after this want to go look at RoboStore.com because I'm now highly interested since we're here talking.

1:26:13Yeah, I'll look at it. I'll look at it. Okay. Unitree H2. Whoa.

1:26:22Let's watch this. Sorry. Can you hear this? Yeah, that was a little in your face.

1:26:38I don't like the face. I wish it didn't put the face on it. Like, in my opinion, it should have no face. The face is a little... First off, it's a little dull. It's also a little haunting. It's uncanny. That it has black eyes is a little uncomfortable. Yeah. There are some examples of... It's got hips on it. That's funny.

1:27:07I'm looking up the robots from my childhood right now on Google and thinking about how far this has come. Well, you really want to see how far it's come. Look at how Doctor Who envisioned. The R1 is for$4 ,900. Yep. Yeah, I thought they had a sub$5 ,000 model now. 40 LIDAR. What about this? Oh, these are all the... Oh, human. Okay. What's this one? Oh, Grant. Yeah. Okay. That's like your RizBot-looking model. Yeah. This one's about$13. RizBot's like a G1. Yeah.

1:27:51If you don't know who RizBot is, go Google tonight. Find him on Instagram. Watch some videos. I'd play him on the channel, but I don't know what I'll get. Yeah. Banned. That's what we'll get. Yeah. Alien Go. What are these quads? What's an affordable quad look like? Was it really$1 ,600? Where does it say that? It said it on the front page on Google when you were looking at it. It showed one of the models as starting at$1 ,600. This one doesn't have a price on it, which implies that it's expensive. Which doesn't imply$1 ,600. Yeah. I say that. they're pretty happy to put a$74 ,000 price tag on a page.

1:28:37So, I mean, you know. Wow. And, I mean, is there like an add to cart feature on these? Can I just literally buy it? Or am I just looking at showcases here? Consumer purchases looks like you can. Okay. What is available for consumer purchases? G1. And the R1. The R1's a$5 ,000 one. Okay, the go-to I think is the Let's see what happens $2 ,800 Okay, that's not unreasonable My problem is, what the hell am I going to do with this dog? Yeah Like, I want it Don't misunderstand me, but my problem is like What do I want it to do? Just scare my pets? I would want this before I want One of those Let me see if I can show you this I want it to be my home security system this guy i like

1:29:37capable of firing billiard balls at people who enter our home in the middle of the night or no for yet llm and robot here i'll look at this one yeah this guy is what i want one of these i just saw somebody on twitter a couple nights ago who built one with their kid and it was really neat yeah it's it's this is super cool i actually love this thing they're not like super functional but they are darn cute. Yeah, I mean, you can talk to it. You can have it do tasks for you. I mean, kind of similar to the things that Flexion was talking about. Like, you know, you can run an agent model that it can talk to and pre-program commands and things like that.

1:30:18Yeah, this is cool. The other one I want to show. And you can program it yourself. Like, you're able to access its code, correct? Yeah. Or at least an app that lets it. Yeah, so actually one of the cool things Actually, I think it was on the robotics channel. Okay, so going back here, let's look at this week. Will it be in the top? It might be. I'm going to get distracted by really cool things. Wall climbing robot. Okay.

1:30:51You know what? It might have been two weeks ago. Let me do top month.

1:31:00oh yeah this is cool have you seen this wheelchair that can just go down regular stairs a whole lot faster than the one you get installed for 80 grand too yeah that's a real use case that's actually awesome it is you wouldn't think it but the agriculture space is full of robotics right now They are all over it. They are using AI for all kinds of things. It's really... He kicked a kid! Is that Riz? Look, he kicked his lady's bag. I'm going to see you first. Oh, this is wild. Look at this. This one's a little more haunting. Yeah. One foot in front of the other. Interesting. Okay. I have lots of problems with that one.

1:31:54Mistral has a robotics team. kind of similar vibe that we saw before with the two hands yeah claws i like this one did you have you seen this one you got the johnny five head again or johnny five ish fauna okay what's the story there really cute um let's pull that up in a second i want to show the the Ricci demo that I'm sure is going to be on here. That's cool.

1:32:26Is she curling with her room? She's curling.

1:32:32Amazing. Oh, wow. This is cool. Okay. Look at that. Okay. I'm sold. I now need the dog because it can carry my groceries. It didn't say that that was an option. I need some kind of, yes, a cart attachment. I need a good mesh basket. Yeah, and some wheels. That's awesome. Yeah, maybe a drink holder for when I have to go get four people coffee. Okay. Yeah, I got use cases. I think we're getting close to the one. Okay, so this is the video that I was talking about. So in this demo that we were going to watch, but the ad was taking too long to play, this guy basically accidentally breaks his reachy.

1:33:24And it's because he's using it wrong. And so the actual creator here explains what's happening. And I'll skim through this shortly. He says basically the fact that this guy couldn't figure it out shows that it's a design problem, which I respect deeply. but then he shows what you're supposed to use so when you turn on this marionette mode basically it goes down and it looks like it's falling asleep um and what's actually happening is it's now a puppet that you can control hence the word marionette so what he does is he goes in and he's kind of explaining that now and he shows that you can move it around and basically what you're doing is you're training it so however you move it in this mode it will then anytime you like play that recording it'll move in that way so he's like moving his hands around and like shaking it and then like swiveling it around like that so then when he goes back and he plays it um the reachy robot will do it now obviously that's like a simple example where it's just shaking his head but it just shows you the type of stuff you can do with this which is really cool I should build one of these for my grandkids for Christmas yeah dude they're really cheap too like 300 bucks right like 299 yeah maybe more like 400 there's like a 300 one and a 400 one again cheap for robots yeah not cheap for bread yeah look at that basically 500 and 300 yeah Yeah, I mean, it's Raspberry Pi built.

1:35:08No kidding. Nine servos. The big one I had, what did he say, 36? I think so, yeah. I think that's the number. He knew the number of servos, but when I asked how many parts, he was like, I don't know. It's a lot. I was guessing in the possibly even, maybe not tens of thousands, but maybe. Yeah. the last one i want to show before you know we can drop off after this if you're down um

1:35:41Nikita Rudin:is that what it's called oh no sunday sunday

1:35:47Nikita Rudin:i think that's it let's see if this is the right thing yeah okay i love this this is my favorite robot that i've seen memo memo yes this is the coolest robot i've seen personally aesthetically especially yeah and if you look at the videos of how they train it um it's like a guy wearing these gloves and he's like doing all these actions it's really cool the uncanny valley vibe is real isn't it yeah but this doesn't trigger it for me a little bit of this is so cool and also Oh, that dash terrifying. Does this trigger it for you? Does this trigger Uncanny Valley for you? Just a smidge. Okay. The hand movements, maybe?

1:36:32I don't know. No, I think it's the eyes. Oh, okay. I think it's that it has those eye things. To me, it has the horror movie vibe to me. I like that it has a little hat. And I like the eyes that make it look like a cartoon, almost like Baymax or something. Now what does Memo cost? Or we don't even know probably I don't think it's out yet They just launched this late last year Beta late 2026 Beta Yeah launching 2026 Beta application Should I apply? I mean I don't see why not Okay Let's apply

1:37:18If you're watching and you would like your robot reviewed. Yeah. We will definitely do this. And brawl over whether it goes to California or Missouri. Yeah. I mean, if it goes to Missouri, I'll go with you. Okay, that's fair. Yeah, to be clear, if it goes to California, I'll come to you. So they built their own foundation model here called the Act One. It's kind of cool. You can swap hat colors.

1:37:54Wow. Fun. All right. Shall we wrap it up? I think that's a good idea. Sorry. I got the yawns this morning, man. See my eyes? Yeah, it's all good. Well, Grant, this has been so much fun. Definitely. Thanks for all the people who hung out till the end. Yes, thank you. I'm glad. I hope somebody watched. We don't see the numbers on our side, so, you know, I know at least four of you were there, by golly. No, no, no, there's a lot of people I've been watching. Oh, have you? Oh, good. I'm glad. I'm glad, because this was a lot of fun. And, you know, hey, physical AI is one of the hottest topics this year.

1:38:35The truth of the matter is we don't touch on it in the normal newsletter so much because it's not necessarily the audience. but to be able to occasionally bring some cool ones on here is very much a thing we can do. And we're always on the lookout. Basically, if it's AI-related, nerdy, and gets one of us a little excited, we'll bring them. Yeah. If they want to come hang out, let's do it. And we're easily excitable. All right. Thanks, everyone, for watching. And if you haven't yet, please definitely subscribe to the channel. and with that farewell for now farewell for now humans I wanted you to say it I couldn't say it I got distracted briefly you have to say it and on that note farewell for now humans we'll see you next time bye bye