Your Personal Open-Source Humanoid Robot for $8,999 — JX Mo, K-Scale Labs

Hello everyone. My name is Jax, and I'm a founding engineer at K-Scale Labs. We build open-source humanoid robots, from hardware to software to, um, machine learning models. And we build it especially for developers. Yeah, so humanoids have been getting a lot of hype recently.

You saw the Tesla Optimus, you have Unitree robots, you have 1X, and etc. They're quite proprietary, and they're quite expensive. And the humanoids are getting so much hype is because of very big problems, like physical labor shortage, consumer household, and also, like, off-world exploration and etc.

Yeah. So for us at K-Scale Labs, our goal is to really solve general-purpose robotics for everyone, and open-sourcing the entire stack to the entire world. So everyone will be benefiting from this really, really useful technology instead of a few different companies.

Yeah, and our team is about 15 people in Palo Alto. You could visit us anytime, we're launching some robots in the next coming month. And, yeah, I'll be demoing some robots in the slides.

Cool. So we're currently working on two robots, the K-Bot and Z-Bot. The K-Bot is a 411 humanoid robot that we made in the last 5 months. It has a full aluminum body, runs an RL controller for local motion, and it has, uh, it's pretty sensor-complete.

So it can do other really cool tasks that the previous presenter was showing. And this will also be one of the cheapest humanoid robots on the market, and it's ready for pre-orderright now, delivered by October. Yeah. So if you come visit, this is a demo we do.

You can kick the robot. The controller is quite robust, and the robot can take a lot of damage. On the robot itself, you also have, basically, you can do VR tele-operation with building hands. I'll describe the modularity a bit more.

But yeah. So the robot's able to run a bunch of different local manipulation policies running our own RL training framework. So when we started building the K-Bot, we thought about, like, how do we make humanoids scale in the future, and what makes it, what makes it possible for people to adopt humanoid robots.

So we basically designed this humanoid robot to be the simplest possible factor as possible, and it's going to be the most affordable development and research-grade humanoid robot at $9,000. The next cheapest option is probably at 40k, which is the Unitree robot.

This is roughly the cost of most robot arms today. Like, if you buy UR5, that's, I think, 15k worth, etc. So the entire robot is going to be open-source. What that means is we're going to have open-bomb. Every single piece of the hardware, CAD design, electronics, PCBs, software, machine learning models will be fully open-sourced, so you can replicate this if you want to.

And it's also going to be very, very modular, which means that you can basically change out end-effectors. We have different mechanical designs that you can easily make interface with our end-effectors, and you can just take out the hand to, for a parallel gripper instead of a five-finger hand, or you can use whatever end-effector you want to use, like the Umi grippers or etc.

And that means we can also easily upgrade and also fix the robot. So our goal with selling this to developers is that when we have new hardware updates, you can easily just re-screw the robot in with brand-new legs or brand-new arms, and also, um, brand-new head.

So, you know, as compute improves, like you get new NVIDIA chips, you can easily just add a new head onto the robot. Yeah, we're also building, we have built the entire Python/Rust SDK for people to use. If you come visit us, you can start programming this robot basically immediately.

It's like a Python-style package, and you can start working on it. And it's capable of running the latest state-of-the-art ML algorithms. So in terms of local motion, you can use, like, NVIDIA Isaacson to train a PPO policy. We use MJX, which I'll explain a bit later, but you can use all kinds of different frameworks.

You can also run, like, different VLMs, like language models, on the robot. Maybe not locally directly, but you can run it through cloud, or you can run VLAs and etc. as well. So this robot, by default, it's going to be 5 DOF arms, but you can easily interchange to a 7 DOF arm.

So that will suit most of the research you need, research labs need. And we'll make continuous, like, model and software improvements with OTA rollouts. So every week, basically, we'll make new changes to the software as we go. Yeah.

Here are some more specs. I can't release the full spec yet because we're launching soon. But, yeah, you see, it uses MIT CHiTA actuators, pretty standard components with, like, MUs, just different audio modules, displays, cameras, and up to 250 TOPS compute currently.

Yeah, and we really started this project in about October last year. So we've been moving pretty fast. We have brought this to mass manufacturing, and we have a new, basically, design that we completed that I can't show you now, but you'll be able to see in 2 weeks.

Yeah. So we started off this, the K-Bot is like the K-Scale Stampede project, like a full-sized humanoid robot that's 3D printable. Then we moved to, like, a prototype, and we worked with different manufacturing partners to actually make the new one that you see.

Yeah, it's launching soon if you're interested in the K-Bot. You can go to https://k.scale.dev/. Yeah. I'll give you a second. No, I'm done, done. Whoa, that's only one robot. I'm not even one-third finished. Joking, joking. Yeah, and so what if you can't spend $9,000 on a cool humanoid robot?

What if your, like, wife or husband doesn't allow you to do it? Well, introducing the Z-Bot, which is a 1.5 feet humanoid robot that we also made at K-Scale Labs. So this started from a hackathon project we did.

It became really popular on Twitter and also WeChat. And so we're bringing this robot to mass manufacturing as well. It runs the same local motion and software stack. Like, that means you can basically program stuff for the small robot, but you can also put it on the big robot.

So, you know, if you make, like, a voice chatting app, you can just put it on either robot, and it runs also a local motion policy as well. It works out without the simulators. Yeah, we really got inspired by the Google DeepMind's robot soccer paper, where, you know, it runs your own play soccer.

That's really how we were envisioning it. So, yeah. We have a pretty good, the launch went really well for the 3D printing one. So our Discord has about 5,000 people. I think a few hundred people have actually made a 3D printed one on the orange one on the bottom.

Yeah. So we also started this project in November, and we are already bringing it to mass manufacturing, which will also be launching very soon. Some people, yeah, we also run, like, monthly hackathons, so you can just come try out the robot.

Yeah. Yeah, same, same website. Okay. Okay, that's the hardware stuff. We talked about the hardware components we just open-sourced. Oh, yeah, also the Z-Bot will be fully open-sourced as well. And so we also open-sourced our entire ML and software stack.

So really, like, our core angle is basically to make this, make the K-Bot autonomous. Well, so, you know, it's a pretty standard dual policy. You have the high-level controller, which is a VLA. Then you have the RL whole-body local motion policy.

Yeah.

So what we really wantright now is to basically finish, we're currently working on both, basically, the RL part and also the VLA part. And we also made our own firmware/software architecture to power these robots in Rust. Yeah. Our angle is basically to make the robot so easy to use.

Any developer could write apps for robots. So, you know, Python application that you can reshare with people. You make the robot do some very specific use cases that can be reused by other people. It's almost like an app store.

And to do that, basically, we offer a lot of really cool developer tools. We've been working on it in the last 6 months. So we open-sourced a library for basically GPU-accelerated robot learning. Well, it's mostly, like, local motion manipulation training.

We used MJX for this. And, yeah, the video of you us kicking the robot, it runs the controller, RL controller, RL model that we trained in this training framework. Yeah, we also are working on to basically being able to integrate and fine-tune all the different VLA and generals policies that you see, from Pi Zero and also NVIDIA Group that we're also presenting today.

So this robot will be able to run, you know, we're trying to make the infrastructure very easy to run any cool models that you see that will be useful.

Yeah, we also made this operating system, which is like a software framework plus, like, a Python interface that you can use to program the robot using Python with Rust. So, you know, you can just, instead of, I don't know, if you guys use Rust 1, Rust 2, they're pretty hard to set up.

But using our system, you can just install a Python package, like pip install kos, and you can start programming the robot. You just connect to IP. It's very, very easy to use. And we also have a digital twin in simulation.

We call it a kos sim. It has the same gRPC interface you can use for controlling the robot in simulation. And all you have to do between programming something in simulation and real is by changing the IP address.

So you can prototype really, really rapidly without having to worry about breaking the robot, which is very cool. So, yeah, this is also fully open-sourced. You can try it today. You can actually program the robot just by using kos sim.

Oh, I don't know what happened to the images. But, yeah. And then what the, what basically what the machine learning and the operating system layers enables for us to run different policies, VLA models, and on our robot hardware, and for people to develop really cool applications with.

So, yeah, I'm just going to go through, like, a very, very quick RL training and deployment examples of how researchers and developers could use our robot to train a local manipulation policy for the robot to, you know, grab different things or walk around or even dance.

So, yeah, RL training setup is very easy. You just get cloned the repository, and then all you have to do is run python-m train. And in this train.py, you effectively have all the training code you need abstracted. It's about 500 lines for walking.

Yeah, and then basically you can run this on, like, you know, using RunPod or your local GPU. It's MJX, so it's like, you know, it's accelerated, accelerated compute. And training a walking policy roughly takes 1 hour to 2 hours.

And, yeah, so you're just going to run through, like, millions of different, not examples, yeah, iterations of the robot performing the task you want, and you can tune the reward functions and etc. And you can see the loss and reward functions in our observability, basically TensorBoard.

Yeah. And afterwards, when the robot's finished training, you can easily evaluate it in kos sim. So all you have to do is, like, K info sim, like, you run the policy, yeah, in simulation, and you see the robot.

If it's walking, if, well, if you can see if it's doing the thing you want it to be doing. For example, like walking, standing, picking up objects. And if that's really good in simulation, then you can just easily change the IP address, and then you have sim-to-real deployment.

Yeah. And it's very cool. Like, you can basically get a robot to work in, like, one-tenth of the time of, like, what you would take to set up most training librariesright now. And also, like, we're a team of 15 people.

We do everything from hardware to software to ML. So how we're able to do this is actually by working with our open-source community. Currently, we have about, like, 5,000 relative active Discord members in a few servers. We have a lot of public open bounties that people are tackling.

And because of our software's MIT licensed, yeah, a lot of people are coming to help us. And we also run hackathons almost on a bimonthly basis that a lot of people come to participate. Yeah. So we're also hiring electrical, firmware, and ML engineers.

So if you're interested, feel free to ask me, and then also go on the https://k.scale.dev/join website. Yeah, we're trying to hire a lot more cracked people to join us.

Yeah, so we're launching the robots and software stack in about 2-3 weeks. If you're interested, follow on the website. I'll be happy to answer any questions.

Yeah. Yeah, sounds good. Go ahead. Yes. Yeah. Where's the power? Where's the power, the battery? Yeah, is it battery packed? Yeah, yeah, it's battery packed. Yeah. It just has a battery. I don't know if I can show you in the picture.

Yeah, it's behind this, basically. You can just slot it in. Like, it clicks in. Yeah. So when the train, you talk about all about the trademark, the trade-offs between the weight of the battery and then what the longevity of these things are.

Yeah, the weight of the battery versus longevity. What do you mean? Yeah, longevity on a chart, like a chart. Oh, like how long it is? Yeah, like how long it is. Yeah. Walking, so far, our test is about 2 hours, but you can pass through.

So you can power through the wall plug. So it's, yeah, you can just keep letting it charge and also run at the same time. Yeah. Yes. Well. Black jacket. Yeah.

What are the use cases that you guys are imagining to start off with? Is this going to be more for, like, commercial, like, you know, factory kind of use cases, or do you envision this more in the home, like, helping out?

Yeah, so basically, our bet, so a lot of companies, like, especially in the US, are betting on, like, B2B. So, like, Figure, for example, are selling to factories. Sims Tesla itself is the customer. For us, our really bet is becoming the first US consumer robotics company.

Like, robot humanoid robotics company. Yeah, so really selling it to anyone that's interested in developing robotics. So a lot of our current customers, we accidentally launched our robots. Like, people accidentally started buying our robots through our Shopify page.

That was a complete mistake. But a lot of people that bought were just people genuinely interested in using for household tasks, like programming it for different research. And also, there are a lot of companies also interested in working with us to make B2B businesses.

Like, for example, the food demo that we just sell. Yeah. Sorry, if I can ask a follow-up question. Yeah, if you can. Like, what household, like, chores do you think it would be well-suited for? Like, unloading the dishwasher, for example.

Yeah, yeah. I mean,right now,right now, we don't have any, I don't think anyone really has a fully working VLA model yet. Soright now, it's pretty limited to tele-operating system, sorry, tele-operation. So, yeah. But soon we hope to be able to have, like, this navigation VLA stack for you to do, like, you know, folding cloth or, like, doing dishwashing as the model capabilities improve.

Yeah. Yes. Oh, white shirt.

Oh, I mean, green shirt. Yeah, you can go first. Can I ask? Yeah, of course. Okay, yeah. So the, you kind of alluded to complexity of Rust 2 in terms of the setup. I'm wondering if there were other, like, benefits and trade-offs that you considered for foregoing something like Rust and Rust 2 in that ecosystem.

Oh, yeah. Like, why not use Rust, basically? Yeah. Well, there are a lot of reasons. Our robot is mostly programmed. So Rust is really good because, like, the nodes and stuff,right? So, like, the async, like, the communication. But for our robot, we really don't have that many sensors.

And we really want to do this, like, model-based, like, policy-based robot. So we don't have many complicated sensors that we need to, like, async communicate at all times. The other part is, like, I'm pretty opinionated. I used Rust 1, Rust 2, Foxy, and Noetic.

I've just had a pretty bad experience using it, having to set up Ubuntu, you know. Like, I just want a robot I can just buy, buy, open the box, it stands where it walks, and then I can just start programming it using my computer.

Yeah. Yeah. What kind of AI accelerator, what kind of AI accelerator is on each of the robots? Yeah, so basically, for the K-Bot, it's going to be Jessen, Neno, and AJX. Yeah. So, yeah, there are different compute options you'll be able to choose when we launch.

Yes, yeah. Go ahead. How do you tele-operate it? So you can just put either VR headset. So there are a few different methods. So the preferred option for a lot of people is VR headset. We have this, like, we also train, like, pseudo IK.

So basically, it's like going to position using, like, an RL model instead of just, like, calculating IK. But it works pretty well with our VR setup. So you can move the hand gestures. You can click button to open and close gripper and just, yeah, move your arm and stuff.

Can you tele-operate the small one too? Yes. Yeah, you'll be able to. It runs the exact same software stack. Yeah. Last question. How do you compare it with the Tesla humanoids? The Tesla humanoids? In terms of, like, mechanical powerness, like, you know, the Tesla is way more powerful.

It has, like, linear actuators and etc. But in terms of, like, actual use cases, I don't think it's really that different. Yeah. I mean, Tesla, it's actually built for, like, a factory type of use cases. But in terms of, like, you want to, for people to actually buy and use this robot, it's not very different.

What's the price difference? I think the last time I heard Tesla Optimus is about 60K, at least. We can ask some Tesla engineers. But our robot is $9,000 before mass production.