Everyone Can Build a Robot: Open Source Embodied AI With Seeed Studio | NVIDIA AI Podcast Ep. 300

Previously, you need to spend months of trainings to understand the spatial planning on the robots, how it moves. But now, what you do with the robots is—after the setting up, you train it, like you're training a dog. You teach it how to do it by holding its hand to do the operations for several times. Then you send all the data back to train on the cloud and deploy on Jetson. Welcome to the NVIDIA AI Podcast. I'm Noah Kravitz. Our guests today are Elaine Wu and Eric Pan from Seeed Studio. Eric is the CEO. Elaine is the head of robotics. And if you're not watching this on video, if you're listening on audio, you might want to check out the video stream. We've got a handful of robot companions with us today. We're going to get some demos. We're going to talk about open robotics, OpenClaw, edge computing, robotics, and AI, all kinds of great stuff. So let's get into it. Eric, Elaine, welcome. Thanks for joining the NVIDIA AI Podcast. Pleasure. Thank you for inviting us. Thank you for having us. And thank you for bringing your friends. So maybe we can start, and Eric, I'll ask you first, for a little bit of an overview for listeners, viewers who might not know Seeed Studio. What the company is and how you moved from open hardware into robotics and edge AI. Sure. So, at Seeed Studio, we have been supporting the global communities on open source hardware from 2008, for 18 years. So we do, starting by open source hardware modules, but we move on to all kinds of devices. We worked with NVIDIA for several years around just nanos, but now we have a lot more robotics, open source robotics, from arms, from little desktop embodiments, so list goes on. So, as NVIDIA's elite partner, I see that our AI robotics team has built a lineup of the Jetson Orin devices as the most powerful and advanced robot brain. Together with that, we also have a full stack of robotic components—from perception and from the control. So, I see that our mission is to make technology more accessible to everyone. And so how is this convergence of AI and OpenAI and robotics, hardware, and technologies now like OpenClaw—how is this shaping the world of robotics and open robotics? I think it's a very evolving and progressive trend. If you're starting to understand microcontrollers, okay, everybody can use microcontrollers. Through all the open source hardware and maker movement. Then gradually, people realized, okay, I have open source hardware, can I use it for industries? Can I do real projects with that? So you see a lot of IoT, a lot of robotics in the early times start to pop up. Because it's easier for everyone, cross boundaries, coming from different industries, now they can use technology for themselves, no matter where they are all over the world. And with all the HAI, all these agentic AI, and people see it differently, okay, I can really get things talking. I can really give a soul to my creations. It's not just a machine. It's a machine that can talk to you, can get assigned on tasks. And this is becoming a new society of robots. And the community, I think, is the root, because you don't need one robot for everything for now. It's actually going to be specialized robots with open source models covering a specific domain. It has to be done by millions of people all over the world. So we are happy to support NVIDIA in this whole process. So along those lines, why is the idea of open so important to robotics? And you hinted at this, and we've been talking with frontier AI companies, you know, building models and building software tools, and just talking about how open source, the open community, is just... things are advancing so much faster than they would be otherwise, with everybody collaborating. Is it a similar thing in the robotics community? Why is open... Seeed Studio has always been built on open. Why is that so important? Yeah, I think a very important factor is people don't know what to build to be the number one creation. So it's not a game. People are trying all kinds of diversities. It's like biology, that you have all kinds of creatures. They compete with each other, they evolve with each other, then we have the best creations for every domain. And through open source, it's not only the technology is evolving fast because people can contribute with their own talents, and also it's a trust. It's for adoption. So imagine you need to have an HAI in your home, in your factories, in your retail stores. You want it controllable. You want to be able to modify it. You don't want it to be locked away. So open source is facilitating a lot of new technologies to push to new boundaries. Instead of salesperson, we have the makers, developers, they naturally, with a curiosity, they're trying to find out what's happening and integrate it into their business. This is, I think, by default, if you're talking about robots, that's the best way. Of course, there are like closed source, the humanoids, there is like traditional companies doing this forever with very expensive ways. But I think what we are looking at is, that's not everything. We at least give the possibility to everyone to create their own physical AI for their belongings. And open also means accessible— accessible to all developers, and everyone can get in touch, for example, physical AI and embodied AI. So I think open, the most important thing is lower the barrier and with the open model, and customizable, and also affordable. And also a bunch of the tutorials that Seeed is keeping them moving, keeping creating that, and to help developers to get, very easily, hands on these different products, hardware, and also the open source dev kit. And also growing together with the ecosystem. I should have maybe asked you this at the beginning, but who are your customers? Who are Seeed's user base? And what are some of the... are there particular domains? You mentioned, you know, manufacturing industry... particular domains that you sell to your products, your robots are being used more than others? We have a... millions of customers, so most of them are calling themselves makers. So it can be everyone— hobbyists, like students, you know but a lot of them are researchers, and they start to understand the possibility of technology and to integrate into their own domain. But they don't stop there. They move on. They become employees. They become startups. They become the project needs in big companies. So you see, like, micro-veins, going into every industry. So it's very hard to define who are our customers, but we see the majority, by number, are the students and researchers. And if we move on, we see a lot of revenue or a lot of productivity coming to small and medium businesses. Because previously, they don't have the power to do that. Now they have all the resources. And they can help each other to very fast adopt technology. And also the big companies, they have problems in innovation. Because the company might be too big, but they have all... they're trying to have their innovation apps. In the innovation apps, how do you find people that can do everything in the fast paces? So the developers, the makers, they very naturally grow into all kinds of domains. What's your best-selling product? Do you sell a kit? How do you, what does... if I'm a maker and I want to get into robotics, what do I buy from Seeed Studio? So for now, we, I think, for robotics, we sell mostly, it's called the SO-ARM. Okay. It's a project we're working with Hugging Face. It's an open source arm, but costing only $200. I've seen that, yes, I've seen this arm. Yeah, yeah, it's super cool. It's a new generation of robots. Previously, you have professional robotists to program it in a very long time, debug it, they have to maintain it. So kind of been very expensive. So you see robots more for medical, for automotive. But robots are more disposable. Yeah, $200. $200, and the kids can use that. How about every business we can have some robots to help us? We did a lot of hackathons last year, and a lot of people are just using it for cooking, for barbecues. It might be like a toy to many professionals, but we have to respect the possibility underneath. If they grow up, if they think seriously, how can they use affordable robots for street business, for a lot of pervasive possibilities? This is actually happening, and I think that's a root, a foundation for more premium robots, premium business. And a lot of people get to learn. A lot of people get to develop this into more mature plannings. And customers can get all of our products at our website. And we also have the global distribution channel. And also, we are pretty active in the developer ecosystem. Like Eric mentioned, we host the hackathons and workshops in different cities and countries. So developers can easily find us. And how easy is it, how quickly could I... I'm not a developer, I'm fairly technical for a content person... how quickly, if I got one of your arms, could I get up and running with it? Previously, you'd need to spend months of trainings to understand the spatial planning on the robots, how it moves. Spatial planning, I'd need an extra couple months, but yeah. But now what you do with the robots is—after the setting up, you train it, like you train a dog. You teach it how to do it by holding its hand to do the operations for several times. Then you send all the data back to train on the cloud and deploy on a Jetson. The Jetson will use a diffusion model to detect with the cameras to try and find the best way to execute the projects. If it's not doing right, retrain it. So it's very, like, organically and naturally. It should be for the skilled person, who is like a chef, who is a blacksmith, trying to teach a robot in an organic way to make it happen. But it's not for someone else. It can be done for themselves. This is my robot. It has, it inherits, or it's my apprentice. Helping me to do something, and it can help me on my business, but not giving away my business. It's not replacing me. It's the robot I acquired, and it's going to enhance my operations. Absolutely, yeah. This is really like, I think it's a different mechanism and different business model. And it's the same way that we talk about, you know, screen-based AI, right? That it's not replacing, it's augmenting, it's... I use AI tools all the time, they help me do things faster, better, you know, but they're not taking my job. So, yeah, it's wild to see this coming into the physical world in, you know, sort of a bigger mainstream kind of way. It's just, it's amazing. I wanted to ask you about, you mentioned, Elaine, the partnership, being an elite partner with NVIDIA. Could you speak a little bit to that relationship and what it brings to the Seeed community? So we started working with NVIDIA back to seven years ago, starting from the Jetson Nano, and then we became the NVIDIA Elite Partner. And then by making the NVIDIA Jetson carrier boards and also the devices. So by this, NVIDIA helped us a lot. So first, I would say first of all, thanks for the trust, for building up this partnership. And then becoming a partner of NVIDIA, we get a very deeply tech support and to help us to get this hardware and the products ready to market. And secondly, I think NVIDIA's emerging technologies is pretty active in their community, like Jetson AI Lab, and also their forum. And also their tech stack, like Isaac SIM, and the Gr00t—from their SDK to their models. So we're pretty actively working together with the ecosystem to put these emerging technologies into our devices, make that as a solution, as the tutorial. Can I ask you... How everybody's talking about OpenClaws... We've got a claw on the table here. How has the integration of OpenClaw onto Jetson accelerated, enhanced your development process? Yeah, so for the OpenClaw, we have the different angle choice. So first, it's application-wise. So now, before, as Eric mentioned, previously, if you want to program a robotic application, that will be very complex. You'll need to program from this perception until it control... I mean, almost every part of the robots, you'll need to hard code that. And now, through the OpenClaw, we have tried... we have installed OpenClaw locally on the Jetson Thor, and also it called the local API of the model, which are the Qwen3.5 certified building model, and then it can do, if I... text on the chat box of the OpenClaw, like move up the robot arm, move down, or pick up the claw, and it can directly execute the task about that. So that is also transforming the way... how we code. It's not coding anymore. It's just the texting to control the arm. And the other way is that it's a way for Jetson users—I think the update, so update the repositories or the update the debug, the issue inside of the Jetson—that will be a time-consuming thing. So if we get the OpenClaw inside the Jetson, it can quickly help you debug or update the necessary thing on the Jetson. So that will be an upgrade in efficiency, yes. And to sum up, I think it's not only just a tool. The robot arm is becoming a claw itself. It's very agentic. So the process is, we connect this about two weeks ago to OpenClaw, and we try to ask it, we give it a wild try. Find yourself the libraries, you are in the SO-ARM, and find your libraries, read the instructions, and put yourself into a physical world. It put together all the libraries and tried to plan how do you... what do you mean by moving 10 centimeters up? So it did all these missions, and we actually gave OpenClaw a physical body. Not only help you to understand the world, but help you to build the world. Help you to move the physical things around you. And the way you master it is, you're talking to a claw through a message, through a WhatsApp, or through a microphone array. So i think this is totally liberating the claw, as sometimes these fields are scary—you know, you just talk to it, and then it starts to move. I was speaking with Harrison Chase from LangChain, and we were talking about agents and this idea of giving an agent an identity, and what that means, right? And you mentioned before about, you know, putting a soul into the robots. Exactly. Right, and so does the robot arm, like... does it communicate to you that it now understands, oh, I have a physical embodiment? Like, is it getting into that kind of thing with the identity of... Exactly. Because you can write in the soul, what's your role, what's your definitions, and what are your skills? So we give, we don't... It's very controllable. We don't ask you to try all the wild gestures. We just give it the actions you want with all the peripherals, with all the objects on the table. Now you are a chef. You should help us to cook this egg. Now he knows, okay, I should be following what kind of SOPs. But I think this is just a start, it's... for a lot of people use OpenClaw as their personal agents, like it'd be part of my job. But I think with all the robots, all the physical AI you put, install OpenClaw into more projects, they are actually far from you. But everybody can talk to them, and they can talk to each other to collaborate. We have a new society of them. But of course, we need to control them in some privileges and their authorities. Right. Yes. But it's a different way we can use agents. And they can have their sub-agents as well. You need a home assistant to manage all the staffs, coordinate them. You maybe need an AI camera to look at everybody working and teach them, okay, you're going to the wrong way. And you can ask the AGV to feed the robot arm something, and the robot arm can operate according to you. Okay, I taught you to do this, now it's for you to repeat it 10 times. I'm imagining in my home, an orchestrator and then a robot to feed the dog. What I really want is a clothes-folding robot. And I know that's very difficult with the articulations and such. But well, we've got these robots here. So this is our very new release, the reBot Arm. So that is also another open source project. Now on GitHub, it's already hit 1.3K stars. And so for this demo, what we're showing is, it's doing the trajectory planning. It is showing it's very smooth. it's very stable. To compare it to the SO-ARM101, it shows it's more robust. And also, we have mirrored this real-to-sim to the NVIDIA Isaac SIM. And so, every position and every actuator is moving. That is, you can see in the Isaac SIM for the simulation. So we give it two missions for the robot. It's called a reBot. So I mean, redefine robots. It's all agentic AI, but it has two... actually now it has three features. First, it's open source, so everybody can contribute. You see some of the parts, they are 3D printed, and we'll release all the files so people can turn this into like a microphone holder for you, to respond to you. They can, according to the purposes, scenarios, they rebuild the robots according to their needs. And the second thing is more applied. You can see it's not designed for learning only. People see it as, and feel like, okay, I want to use it. You can do something with it, yeah. And then, the obstacle is, how much... would it be too expensive for my cooking, for my small business? No, it's just $1,000 to start with. This one's $1,000. $1,000. It will be less than $1,000. And we've paired it with Jetson Nano to do the local functions. You don't need a token. You buy a Jetson Nano. Everything is running on it. And you can work with, of course, the cloud or bigger, like Thor, to do more complex plannings. So you can have a management system with more HAI, but this more AI will be executing the SOPs accordingly. So we hope this can enable open source applied and agentic usages. Amazing. Do you, you know, you mentioned, obviously, that we want to stay as humans in control of the system, right, of the robots. And when you're working with tools like OpenClaw, and actually... I'm wondering, can OpenClaw write skills for robots? Sure. It extends, it doesn't matter, it's, yeah. Exactly. So are there concerns around safety issues? Are there particular guidelines that are specific to the robotics domain when you're working with AI and thinking about AGI? No. Because we just started to use the claw to do that. We're still trying on the limits, but for now, I think we... I forget, this just came out like a month ago. Exactly. You wake up every day and you, okay, where are we now? The terrain has been changing too fast. Every day. But what we are trying to start with is, we see as a robot first. So it has all the robot guidelines we should follow. And if we connect to IT system, it should really have all the controls as if it's a human being. So we apply the basic layer first. Then we'll explore the boundaries, and we have a panic button we can just shut it off. It's easier, because you don't need to unplug your wires for laptop, you just unplug it, and it shouldn't start. So thinking kind of, I don't know, a little bit abstractly and kind of thinking ahead, you know, with so many people using OpenClaw, and Claw Code, and these agentic tools, and LangSmith, and creating all of these just thousands, millions of digital creations, you know, every day being built by these agents, how do you think about applying that mentality and applying these digital creations and kind of manifesting them in the physical world through robotics? I think it's like following the OpenClaw discussion, OpenClaw is becoming an interlink between the digital world to the physical world, and now we can control robot arms. But like a lot of details, how do you really move the parts? How do you really interact with your environment? So I think that's where Isaac SIM is coming into play. It's mirroring all the modelings, and you can simulate before you acquire the robots, before you deploy them. And they get very precise renderings afterwards, so you can have the digital twin between them. And a lot more people before, they don't have the resources. I think until now, we don't have one robot for everyone. So how about we, more people can participate first through the online, through the simulations. But we built more sim-to-real, real-to-sim bridges. And as we have more cheap, affordable robots. The sim-to-real gap will be very easily closed because a lot of people validating, a lot of people are improving on the controls, on the details to make them smooth and practical. And then you train the VLAs on these simulations and deploy in the field, they can very organically, for most of the people, they use them as if they are talking to a... a creature. Yeah. This is going to be very exciting. It's amazing. It's amazing. So Elaine, can you talk about how open robotics, open frameworks, open models sort of help developers kind of bridge that gap, bringing their ideas into the physical world through robotics? Yeah, the first I want to point out is the Hugging Face LeRobot framework. So that is the transfer means, or way how we're building the robots. So as we mentioned, the OpenClaw guide is very easy. You can control the robot arms through the text, the chat box. But using the others, LeRobot, there's a lot of researchers and AI engineers doing the algorithms. Or the hardware engineers, they can get touch into the embodied AI, no matter what their background is. Because LeRobot really works because they provide the Hugging Face, they provide it with the datasets, with the model, and with also the policy, and into the one set of the framework to train the robot. That is end-to-end learning, rather than you need to code each part of the robot. Yeah so, and also, I see that we focus on the hardware. So together with the LeRobot, we bring the SO-ARM, we also have our reBot arm, compatible with LeRobot, to help the developers to get easily quick-started. And also, besides our hardware, I see that we also provide the services, the customer services, and also manufacturing services. How quickly we help the startups to get into that. For example, we work with Hugging Face for this Reachy Mini. So we quickly, in five months, from the design to developer and manufacture that. In five months, we started 3,000 units shipped to every customer. That's amazing. Wow, that's so fast. So what's next? What's on Seeed Studio's agenda for the year? Are there different types of robots? Are there different frameworks? Or is it all-in on OpenClaw? What can you tell us? Yeah, I think what we created is a beginning, and what we offer is a reference design. Because we're an open source company, why we ever rebuild the robots open source. So we don't... have expectations that we do something everybody will just use. They will come back to us, they always—can I change this a little bit for my creations? I'm doing a startup, can I base on your creations and to wrap it up with my practices? Why not? So one key part of our business is we help people to scale. So we give people reference design as a framework. They do the development, they come back to us, okay, let's change and make it into some different creations. So we hope to see more of this happening. They have more physical AI, they look at open source design. They have a problem to resolve. Now they have much faster, better, easier solution. And that's where we want to facilitate. We hope we can foster and support a big fleet of physical AI creations. I think that will be very fast to happen in the next one or three years. So you're able to customize the physical design of the robot to meet the customer's needs. Are you building, are you getting into, I don't know, like humanoids or quadrupeds, bipeds, that kind of thing, or is it mostly arms? We actually do it the reverse way as humanoids. We don't want to be the one robot that does everything. We don't want to build a general robot. But we disassemble the humanoids. We build the head and the torso, we build the arms, and we build the wheels. So we start building more parts of the robots. and then people can combine them, according to their usages. So very organically, they can find their practical loop, flywheel, of physical demands and the possible hardware. Yeah, and also for our product line, we have a lineup of the LeRobot-compatible robots. We have arms, we have chassis, we have hands, and also desktop robots. Oh, fantastic. Amazing. A very important trend is that we don't wait for someone to build a humanoid for us. Maybe we don't trust it, maybe we cannot wait, maybe it's too expensive. But now everyone can build a robot for themselves. And if they have a cluster of needs, a community with similar scenarios, someone can create a business on the physical AI. I think that's more organic. It's amazing. There's something about the physical world for all that we're talking and using and developing these on-screen tools. Seeing it embodied in the physical world just is a mind shift, and it's just amazing to see. Eric, Elaine, for viewers, listeners who want to learn more— maybe they want to find out about getting some robot parts of their own. Seeed Studio. Is it SeeedStudio.com? Yep. And with three E's. Strictly with the three E's. Okay. And other places they can go learn more. Social media... We are on most of the social media—Twitter, Facebook, Instagram. So just search us and you'll find us. Perfect. Well, again, Eric and Elaine, this has been fascinating, and as you said, it's just the beginning. So, looking around at what's on the table now makes me think what might be on the table next year. Appreciate you coming. Thank you so much for joining the podcast, and obviously, all the best with everything Seeed's doing. Very nice talking, and I'm proud to share what we have been doing. That's great. Thank you for having us. Our pleasure.