This Robot Startup Is Coming for Human Labor

Jordan Kesslow, Goanna Capital. By 2030, there will be more humanoid robots than there are people in New York City. Someone has to build them. We came to London to meet the team actually doing it. Over the past 150 years, a few inventions changed the world. Phones, cars, computers. And the next big thing will be humanoids. I started Humanoid to free people from dangerous, monotonous, repetitive tasks by building most reliable, safe, and helpful humanoid robots. I'm Artem Sokolov, founder and CEO of Humanoid. Artem, how are you? Hey, man. Nice to meet you. Yeah, great to see you. Thanks for having us. So we're here with Artem Sokolov, founder of Humanoid. We're going to get a little tour of the facility and learn more about the company. So for those who don't know, tell us, what is Humanoid? Humanoid is a company who is building humanoid robots. And our goal is to be the number one player in the humanoid industrial space. And we are the youngest company in the field, but we are the fastest. And we are going very fast. It's already come a long way. I got a little briefing before we came in. It's incredible. I can't wait to show everybody what you have going on here. Yeah, let's go. It's 10 o'clock. It's a bit early for our engineers, but let's see what we have. Cool. Yeah, show us around. Let's see what you got. Yeah, we moved here one year ago. We have here around 160 people in our London office. Yes. We have also offices in Boston and Vancouver. 30 people in Boston, 30 people in Vancouver. Boston and Vancouver, it's more about hardware. And we have more hardware engineers there. And we have a lot of people from Boston Dynamics. We got the best engineers from BD. In our Boston office, we got the best engineers from Century AI in our Vancouver office. And we have more people from Tesla Optimus, One X, Agility. Yeah, more than 50 veterans from different humanoid leaning companies. Very experienced team. What's something that you look for that's not on a resume when you're hiring? What are some of the attributes that you look for? Energy, speed, and ambitious. Three main points that we need to see. That's always a winning formula. Yeah, for sure. You hear that? We need energy, people. Energy. Here you can see our head. It's very interesting and complex head. We have six RGB cameras. We have two depth sensors. Picture eye screen. You'll see working robot later. So the six cameras, that gives you a 360 view? 360 degrees, yeah. For 360 reasoning capability. Gotcha. And is there some sort of LIDAR capability in there? We have cameras on the head and two LIDARs on the base. We are using both. This is our wheeled robot. Our immediate focus for industrial application. This is cool. So this can roll around? Yes, and it's also omnidirectional. And this is our main focus for this moment. Because we started with both versions. Bipedal one and wheeled. And in parallel, we've spoken with more than 150 clients. All of them from top 500 Fortune list. And all of them ask us to provide them a wheeled robot. Because it's already safe. It's already stable. And already deployable. Because you can use the current AMR track for regulation. Makes sense. So you don't have any possibility of tipping over. Yeah. And also, it's more cost and energy efficient compared to bipedal robots. Okay. And also, you can see that we use parallel joint grippers. It's a two-finger. And we've developed three-finger. And also, we chose this because it's more reliable and cost efficient compared to five-finger hand. Because we analyzed maybe more than 1,500 different use cases with the clients. And we recognized that we can cover more than 90% of these use cases with wheels and parallel grippers. We might as well just use that. And we simplified this solution as much as possible to be ready for mass production and meet our cost targets. Got it. To have a win-win strategy with the clients. Because we offer two revenue options for our clients. One is the ROAS subscription model. It's $5,000 per month. Okay. Everything is included. Okay. And with this model, clients get a payback period less than one year. Fantastic. So you don't have to pay for the whole robot. You can just do the $5,000 a month. Yes. And with $5,000 a month, it will cost you less than 50% of the equivalent human labor per hour. Because robots will work 24-7. Robots don't need to eat, rest, or whatever. Of course. And is there an option to buy? How much is it if you want to purchase? Yeah. We will start just with the subscription. Okay. Just with the subscription to start. For the first two years. Got it. Because we want to be more flexible on fleet management, service, maintenance, and et cetera. But in two years, we will also add the CAPEX model. It will be $70,000 per robot plus 15% subscription. Every year, you need to pay 15% on top of it to update the AI models. Because with the same hardware, productivity will increase every month. When we deploy more robots, there will be much more data, better training, faster manipulation, faster productivity, better ROI for the clients. Makes sense. And you're making these 100% here in London? Yes. Fantastic. For this moment, we will build 17 robots in-house. We will continue building 100 of them until Q1 next year. And what's also interesting, we just signed this week with Bosch. I saw that. Yeah. And we already started work with them to do all optimization and preparation for mass production. And we will start mass production with them in Germany next year. And that's with Schaeffler, right? Bosch will do assembling and manufacturing. And we will use also our own design actuators, which we will produce with Schaeffler also in Europe. Got it. All strategic components we just source or produce in Europe. Congrats on the partnership. That's a good one. Yeah. And it's really like Bosch is one of the best manufacturers. And Schaeffler is really one of the best partners on the actuation side. Because we have around 30 actuators per robot. And actuation is 50% of robot cost and complexity. Is that right? Yeah, it's one of the most difficult parts of the robot. So half the cost are the actuators. Yes. Incredible. And also a big advantage on wheeled robots that we can put bigger batteries on base compared to bipedal robots. And that just probably helps with stability as well as longevity. All right. What do we have here? Some testing happening? Yeah. Here like a more lab space. We do different testing here. We do also, we build physical twins. Our customers provide to us their equipments for different use cases. Boxes, shelves, materials, like different SKUs to train AI models and collect data before we'll go to their real facilities. Okay. Let me ask you about that. So someone's interested in taking your robots. First you find the use case and you test it first to make sure it's going to be practical for them. Is that right? We already know that it will be practical for them. We just need to collect a bit more data to train our AI models. Got it. Because first touch point with them, we're analyzing what they have in terms of priorities. So what types of use cases, where is demand, what is the complexity of use cases and what is their priority, first of all. After that, we have a priority list of different use cases and we just started one by one. So what is this one being tested for right now? It's kind of tools handling. You need to put on the shelf, from the shelf, working with different tools, move them around and et cetera. We started working with warehouse, logistics, FMCG, wholesale. After that, we added everything around automotive and manufacturing. I see some action over here. Still R&D platform. We spend maybe here 10-15% of our time because now our full focus is on wheeled robots. But we continue experimenting and build capability. Because this is our mid and long term strategy for service and household. And you can see that it's a really big, big robot. Wow, it's got a lot of weight. 100 kilograms. 100 kilograms? Yeah. Wow. And is this just for safety, this harness? Yeah. Got it. And also you can see that we have also five finger hands and we can swap easily from grippers to hands. And it's more like the status and fine manipulation is needed. Got it. So what exactly are they testing here? They are working on locomotion, safe locomotion. They are testing different behavior. And we are training our reinforcement-learned body controllers to do very stable walking. Amazing. And he's controlling it with that? Yeah. It's a full operation, yeah. We don't spend now any time for AI autonomous function on biped, bipedal. Like fully focused on industrial. Do the guys below hear all this? Fortunately, no people there. So what do we have here? What's interesting, for example, I mentioned Schaeffler, our client, our partner. And I mentioned that we are building physical twins in our lab to train the models. For example, this is rings from the Schaeffler production. Is it like a bearing? Yeah, bearings. Because they are producing a lot of components for automotive. And this is one of them. They have around 1,000 different sizes of these rings. And they have a lot of stations. And it's very difficult to automate this use case just with stationary arms or cobots. You need to have humanoids. Because you need to move from one station to another station during the day. And it's best fit for humanoids to automate this type of workflow. I see. So they're going to be sorting this in some sort of way? Speaking, sorting, moving around. Got it. Also packaging after the assembly. Sure. What do we have here? Yeah. We spend a lot of time more on industrial space. And this is our example for more service and retail. It looks like a concession stand. We call this RoboStore. RoboStore. Yeah. And now we are talking with different big retailers to do fully automated stores across Europe. Okay. So the idea is you can come up to this robot and essentially order something. And it'll hand it to you, take your money. It's like a kiosk with robots on streets, on shopping malls, and so on. I like it. It's not our primary focus, but we are experimenting with retail front operations also. Let's order something. So this works? We can test this out? Let's try. We have Mongo. And now you have an integrated LLM, right? You can speak to this? Yeah. We have LLMs, we have VLMs, and we have VLA here. We've developed, it's called Kinetic. It's our AI architecture that consists from VLAs. For those who don't know, what is VLA? Vision Language Action Model. Okay. It helps robots do navigation manipulation and so on. Also, we have VLMs, Visual Language Model, that help robots see and do reasoning and understand what is going on. And we have our own reinforced non-body controllers, locomotion for all robot joints. And we have System 3. It's our agentic fleet orchestration layer that helps robots work together, communicate together, to optimize how use cases work for maximum efficient way. Got it. Because we don't want to integrate a single robot. We want to integrate a fleet of robots with 50 to 200 robots per facility. And this is very important to have this orchestration layer to maximize maximum effect from a fleet of robots. Got it. Just to give you an understanding of what you're seeing here, this is what Arthur mentioned as our System 3, the fleet coordination layer. And you can see a visualization of that just for this application. So, this can be used for industrial or commercial settings. Here, it knows more or less where all the objects are. So, you can see that we have dried mango, or the water, or the love corn in the front. And then you have some Pringles crisps in the back. So, it also knows which robots are available and what the status is, following orders it has completed. So, this is what connects to the business layer or the facility management layer. It takes the orders and then distributes them to the robots according to their skills, to their status, or what, let's say, position they have in a specific facility. So, you can go ahead and try it. You can long press this button. You can order up to two items. And then you'll be interacting with our System 3. Okay. In real time, you can use, let's say, any language you want. So, you can give it a go in any kind of language. Okay. Here we go. Ready? This is awesome. May I please have some Pringles and water? I understood. Pringles and water. Please confirm or cancel. Press it and say confirm. Confirm. So, you see your order is being processed. The system is distributing the orders to the robots. So, you can go there. Oh, look at this. Oh, this is sick. Oh, it holds on really tight. You have to wait for it to let go. Yeah, yeah, yeah. It's identifying if there's somebody in front trying to reach for the object. Otherwise, it will not let it go. Amazing. All right. So, I got my water. Now, it's Pringles time. That was pretty good. The robot understood that there are no Pringles. Oh, okay. So, it's got to go get them. Yeah. Pick from the shelf. That was awesome. It's still a little bit slow because we slowed down in terms of safety. And now, we are working more for safety layer. But it will be much faster. I like the rollers. I mean, walking around is not helpful. In a lot of these use cases, you might as well just have them on the roller. I totally get it. This is awesome. So, is it giving it to me? Oh, no. It's going to hand it to me. Oh, here we go. Also, it's also a shelf gripper. And we have a problem that we don't have a touch force feedback. But in our grippers, we already have. Wow, it's pretty strong. Yeah. So cool. Well done. That was successful. Should I put it back in here? You can take it. Well done. That's fantastic. Wow. Success. We'll show it later. Later. Okay. Thank you, guys. That was awesome. Wow. It never ends. You got a big place here. So, this is more testing? Yeah, like a testing data collection. Also, we run real here in person learning real time. Here you can see how we collect data through robot teleoperation. It's one of our two data collections. Humans collect a lot of data to do teleoperation. This is the most efficient and qualified method of collection. Here you can see inside robots. We have around 600 unique items in one robot. Very complex hardware piece. So, what's being trained here? We collect a lot of different manipulation tasks to pre-train our own VLA because we need very diverse data, different style of data, different objects. So, is this a matter of just doing many, many episodes of the same thing so it learns? Yeah. So, it's the same iterations. It just learns more and more. With different objects. Got it. To do boxes, objects, like different types of... And also... Yeah, what is the foot pedal right here? It's just to fix the episodes. One episode, second episode. Oh, it's like a reset. Yeah, yeah, yeah. Got it. Very cool. So, you just have the hands here. Yeah. As I mentioned, we have these devices, the same grippers, and humans collect much more data for us. We do it internally, we do it externally, and also in the next couple of months we will provide for our clients these devices to collect data in their facilities. And also we are developing an egocentric setup. It will be some kind of suit that we will provide for our clients, for their workers, and they will wear cameras, one stuff with gloves, another one without gloves, and their customers, workers, will collect much more data. So, that's to be able to train for specifically what they're doing. Yeah, and I'll go to collect the biggest industrial data set to build the best vision language action model. So, you could essentially have them working for months and just collect a massive amount of data. They will do their usual work, but at the same time collect data. Oh, that's great. And this data will be most useful and qualified because it's not like somewhere in India or China, somebody collects some kind of data. It's exactly what we need. You can probably get thousands of episodes. It's interesting that before we started to develop our own hardware, we bought all available human noise from China, from Korea, and we tested them because we thought that we can just build software, AI software integration application layer. Using third-party hardware. What made you decide to make your own hardware? We tested a lot and we tried to do real work in real facilities, but we fucked up with over-the-shelf hardware, and we decided to do our own for speed reasons, for time-to-market reasons. My mind is blown. I know you see this stuff every day, but from an outside perspective, it's really like, wow. Starting point, starting point. Yeah, we have 17 robots this year. We'll have 100 robots. It will be more fun. How many robots are deployed at the moment? We don't do full deployments now because we have, like, not enough robots. So we do two, three weeks of pilots with the clients. Okay. Signing contracts and preparing for delivery next year. Got it. Let me ask you this, if you don't mind getting into it. You've grown so quickly in two years. Who have been some of your early investors, and how many rounds have you had so far? Yeah, it's just a company fully self-funded by me for this moment. You're kidding. I've invested $120 million from my previous exits in the company for speed reasons. So you own 100% of the company? And stock options for my team, of course. I mean, where do you see this going? I know even next year will be amazing. Let's just say in five years. Where do you see this scaling to? What kind of capabilities will there be? What do you see the outlook there? Yeah, in three years, we want to deploy hundreds of thousands of robots. In five to six years, it will be millions. And in 10 years, it will be hundreds of millions. And when we do these deployments, we will see a world of abundance. And our big goal is to eliminate these dangerous, repetitive, boring, and hard tasks. And also make work optional in the future. And also eliminate hunger, even loneliness. When we do household robots, it will be like your best friend. Also, I can't wait for this future. Well done. I'm excited for you. Artem, thank you for your time. This was a pleasure. Guys, stay tuned for Humanoid. Exciting things happening. Thanks for joining us. Remember to like and subscribe for more videos. The robots aren't coming. They're already here. The only question is what we build them to do.