Your Personal Open-Source Humanoid Robot for $8,999 — JX Mo, K-Scale Labs
00:00:00.000 | JXX LIU:
00:00:15.000 | JXX LIU: Hello, everyone.
00:00:15.600 | My name is JXX.
00:00:16.480 | And I'm a founding engineer at K-Skill Labs.
00:00:19.280 | We build open source human robots from hardware
00:00:22.080 | to software to machine learning models.
00:00:26.520 | And we build it especially for developers.
00:00:29.360 | JXX LIU: Yeah, so humanoids have been getting a lot of hype
00:00:31.920 | recently.
00:00:32.620 | You saw the Tesla Optimus.
00:00:33.980 | You have Unitree robots.
00:00:35.140 | You have 1x and et cetera.
00:00:38.300 | They're quite proprietary, and they're quite expensive.
00:00:40.860 | And the humanoids are getting so much hype
00:00:42.820 | is because of very big problems, like physical labor shortage,
00:00:46.700 | consumer household, and also like off-world exploration
00:00:51.160 | and et cetera.
00:00:52.900 | Yeah, so for us at K-Skill Labs, our goal
00:00:55.540 | is to really solve general purpose robotics for everyone,
00:00:58.720 | and open sourcing the entire stack to the entire world.
00:01:02.240 | So everyone will be benefiting from this really, really
00:01:04.760 | useful technology instead of a few different companies.
00:01:10.900 | Yeah, and our team is about 15 people in Palo Alto.
00:01:14.740 | You could visit us anytime.
00:01:15.900 | We're launching some robots in the next coming month.
00:01:19.020 | And yeah, I'll be demoing some robots in the slides.
00:01:22.560 | Cool.
00:01:25.940 | So we're currently working on two robots, the K-Bot and Z-Bot.
00:01:29.560 | The K-Bot is a 4.11 humanoid robot
00:01:33.180 | that we made in the last five months.
00:01:35.060 | It has a full aluminum body, runs an RL controller
00:01:37.340 | for locomotion.
00:01:38.620 | And it's pretty sensor complete.
00:01:41.140 | So you can do all the really cool tasks
00:01:43.140 | that the previous presenter was showing.
00:01:46.300 | And this will also be one of the cheapest humanoid robots
00:01:48.760 | on the market.
00:01:49.640 | And it's ready for pre-order right now, delivered by October.
00:01:54.260 | Yeah.
00:01:54.920 | So if you come visit, there's a demo we do.
00:01:58.900 | You can kick the robot.
00:02:00.800 | The controller is quite robust.
00:02:02.520 | And the robot can take a lot of damage.
00:02:06.140 | On the robot itself, you also have--
00:02:08.240 | basically, you can do VR teleoperation with building hands.
00:02:12.080 | I'll describe the modularity a bit more.
00:02:14.760 | But yeah.
00:02:16.980 | So the robot is able to run a bunch
00:02:19.040 | of different local manipulation policies,
00:02:20.900 | running our own RL training framework.
00:02:24.760 | So when we started building the K-Bot,
00:02:26.760 | we thought about how do we make humanoids scale in the future?
00:02:30.880 | And what makes it possible for people
00:02:34.420 | to adopt humanoid robots?
00:02:36.080 | So we basically designed this humanoid robot
00:02:38.920 | to be the simplest factor as possible.
00:02:42.460 | And it's going to be the most affordable developer
00:02:45.240 | and research-grade humanoid robot at $9,000.
00:02:48.880 | The next cheapest option is probably at $40,000,
00:02:52.040 | which is the unit tree robot.
00:02:53.920 | This is roughly the cost of most robot arms today.
00:02:57.800 | Like if you buy UR5, that's, I think, $15,000, et cetera.
00:03:02.320 | So the entire robot is going to be open source.
00:03:04.800 | What that means is we're going to have open bomb.
00:03:07.160 | Every single piece of the hardware, CAD design, electronics,
00:03:10.100 | PCB, software, machine learning models will be fully open sourced.
00:03:14.160 | So you can replicate those if you want to.
00:03:17.920 | And it's also going to be very, very modular, which means
00:03:20.700 | that you can basically change out end effectors.
00:03:23.880 | We have different mechanical designs that you can easily
00:03:27.040 | make interface with our end effectors.
00:03:29.680 | And you can just take out the hand to--
00:03:32.040 | for a parallel gripper instead of a five fingers hand.
00:03:35.920 | or you can use whatever end effector you want to use,
00:03:39.520 | like the wumi grippers or et cetera.
00:03:41.000 | And that means we can also easily upgrade and also fix the robot.
00:03:44.800 | So our goal with selling this to developers is that when we have new hardware updates,
00:03:49.200 | you can easily just re-screw the robot in with brand new legs or brand new arms
00:03:53.800 | and also brand new head.
00:03:56.760 | So, you know, compute improves, like you get new NVIDIA chips,
00:04:00.440 | you can easily just add a new head onto the robot.
00:04:03.240 | Yeah, we're also building--we have built an entire Python/Rust SDK for people to use.
00:04:11.120 | If you come visit us, you can start programming this robot basically immediately.
00:04:14.920 | It's like pip install package and you can start working on it.
00:04:17.880 | And it's capable of running the latest state-of-the-art ML algorithms.
00:04:22.840 | So in terms of local motion, you can use like NVIDIA, Isaacson to train a PPO policy.
00:04:28.680 | We use MJX, which I'll explain a bit later, but you can use all kinds of different frameworks.
00:04:34.160 | You can also run like different VLMs, like language models on the robot.
00:04:37.760 | Maybe not locally directly, but you can run it through cloud.
00:04:40.600 | Or you can run VLAs and et cetera as well.
00:04:45.160 | So this robot, by default, it's going to be five DOF arms,
00:04:48.960 | but you can easily interchange to a seven DOF arm.
00:04:51.880 | So that'll suit most of the research need, research labs need.
00:04:55.400 | And we'll make continuous like model and software improvements with OTA rollouts.
00:05:00.680 | So every week, basically, we'll make new changes to the software as we go.
00:05:05.160 | Yeah, here are some more specs.
00:05:09.240 | I can't release the full spec yet because we're launching soon.
00:05:11.720 | But yeah, you see, it uses MIT cheetah actuators, pretty standard components with like MUs,
00:05:20.520 | just different audio modules, displays, cameras, and up to 250 TOPS compute currently.
00:05:30.680 | Yeah, and we really started this project in about October last year.
00:05:35.240 | So we'll be moving pretty fast.
00:05:36.680 | We have brought this to mass manufacturing.
00:05:39.480 | And we have a new, basically, design that we completed that I can't show you now,
00:05:43.800 | but you'll be able to see in two weeks.
00:05:45.800 | Yeah, so we started off this -- the KBOT is like the KScale Stompy project,
00:05:50.600 | like a full-sized humanoid robot that's 3D printable.
00:05:53.480 | Then we move to like a prototype, and we work with different manufacturing partners to actually make
00:05:58.520 | the new one that you see.
00:06:00.840 | Yeah, it's launching soon if you're interested in the KBOT.
00:06:06.920 | And you can go to kscale.dev.
00:06:08.360 | Yeah, I'll give you a second.
00:06:11.320 | No, I'm done.
00:06:12.520 | Whoa, that's only one robot.
00:06:14.200 | Not even one-third finished.
00:06:16.040 | I'm joking, joking.
00:06:17.960 | Yeah, and so what if you can't spend $9,000 on a cool humanoid robot?
00:06:22.680 | What if your, like, wife or husband doesn't allow you to do it?
00:06:26.520 | Well, introducing the ZBOT, which is a 1.5 feet humanoid robot that we also made at KScale Labs.
00:06:36.280 | So this started from a hackathon project we did.
00:06:38.600 | It became really popular on Twitter and also WeChat.
00:06:42.120 | And so we're bringing this robot to mass manufacturing as well.
00:06:47.000 | It runs the same locomotion and software stack.
00:06:49.480 | Like, that means you can basically program stuff for the small robot,
00:06:54.360 | but you can also put it on the big robot.
00:06:56.840 | So, you know, if you make, like, a voice chatting app,
00:06:58.760 | you can just put it on the idle robots.
00:07:01.080 | And it runs also locomotion policy as well.
00:07:03.880 | It works out with all the simulators.
00:07:06.200 | Yeah, we really got inspired by the Google DeepMinds robot soccer paper,
00:07:12.680 | where, you know, it runs around play soccer.
00:07:14.440 | That's really how we were envisioning it.
00:07:16.600 | So, yeah, we have a pretty good-- the launch went really well for the 3D printing one.
00:07:22.920 | So our Discord has about 5,000 people.
00:07:25.400 | I think a few hundred people have actually made a 3D printed one on the orange one on the bottom.
00:07:31.400 | Yeah, so we also started this project in November, and we are already bringing it into mass manufacturing,
00:07:39.320 | which we'll also be launching very soon.
00:07:41.160 | Some people-- yeah, we also run, like, monthly hackathons, so you can just come try out the robot.
00:07:46.360 | Yeah, same, same website.
00:07:50.760 | Okay.
00:07:51.800 | Okay, that's the hardware stuff.
00:07:53.800 | We talked about the hardware components we just open sourced.
00:07:57.800 | Oh, yeah, also Zbot will be fully open sourced as well.
00:08:00.840 | And so we also open sourced our entire ML and software stack.
00:08:05.960 | So really, like, our core angle is basically to make this-- make the kbot autonomous.
00:08:13.320 | Well, so, you know, it's a pretty standard dual policy.
00:08:17.080 | You have the high-level controller, which is a VLA.
00:08:19.320 | Then you have the RL whole body locomotion policy.
00:08:22.040 | Yeah.
00:08:25.400 | So what we really want right now is to basically finish-- we're currently working on both, basically.
00:08:30.520 | The RL part and also the VLA part.
00:08:32.760 | And we also made our own firmware/software architecture to power these robots in Rust.
00:08:37.240 | Yeah.
00:08:39.480 | Our end goal is basically to make the robot so easy to use.
00:08:41.960 | Any developer could write apps for robots.
00:08:46.200 | So, you know, Python application that you can re-share with people.
00:08:49.240 | You make the robot do some very specific use cases that can be reused by other people.
00:08:54.360 | It's almost like an app store.
00:08:56.200 | And to do that, basically, we offer a lot of really cool developer tools we've been working
00:09:01.960 | on in the last six months.
00:09:03.160 | So we open sourced the library for, basically, GPU-accelerated robot learning.
00:09:08.440 | Well, it's mostly like locomotion manipulation training.
00:09:11.480 | We used MJX for this.
00:09:13.720 | And, yeah, the video of you saw us kicking the robot, it runs the controller--
00:09:18.840 | RL controller-- RL model that we trained in this training framework.
00:09:23.240 | Yeah.
00:09:25.480 | We also are working on to basically being able to integrate and fine-tune all the different VLA
00:09:31.240 | and generalist policies that you see from Pi Zero and also NVIDIA group
00:09:34.680 | that we're also presenting today.
00:09:36.680 | So this robot will be able to run--
00:09:38.520 | you know, we're trying to make infrastructure very easy to run any cool models that you see
00:09:42.760 | that will be useful.
00:09:43.560 | Yeah.
00:09:47.240 | We also made this operating system, which is like a software framework plus like a Python
00:09:52.040 | interface that you can use to program the robot using Python or Rust.
00:09:55.880 | So, you know, you can just instead of--
00:09:57.800 | I don't know, if you guys use Rust 1, Rust 2, they're pretty hard to set up.
00:10:01.080 | But using our system, you can just install Python package, like pip install KOS,
00:10:05.000 | and you can start programming a robot.
00:10:06.760 | You just connect to IP.
00:10:08.120 | It's very, very easy to use.
00:10:09.960 | And we also have a digital twin in simulation.
00:10:13.000 | We call it a KOS-SIM.
00:10:14.120 | It has the same gRPC interface you can use for controlling robot in simulation.
00:10:18.920 | And all you have to do between programming something in simulation and real is by changing the IP address.
00:10:24.680 | So you can prototype really, really rapidly without having to worry about breaking the robot,
00:10:28.920 | which is very cool.
00:10:30.120 | So, yeah, this is also fully open sourced.
00:10:32.760 | You can try today.
00:10:33.960 | You can actually program the robot just by using KOS-SIM.
00:10:38.200 | Oh, I don't know what happened to the images.
00:10:43.480 | But yeah.
00:10:44.040 | And then, basically, what the machine learning and the operating system layers enables for us
00:10:49.720 | to run different policies, VLA models, and on our robot hardware.
00:10:54.280 | And for people to develop really cool applications with.
00:10:56.680 | So, yeah.
00:10:58.840 | I'm just going to go through like a very, very quick RL training and deployment examples
00:11:02.840 | of how researchers and developers could use our robot to train a local manipulation policy
00:11:09.080 | for the robot to, you know, grab different things or walk around or even dance.
00:11:15.320 | So, yeah.
00:11:15.880 | So, yeah.
00:11:15.880 | Our training setup is very easy.
00:11:17.480 | You just get cloned the repository.
00:11:19.560 | And then, all you have to do is run python-m train.
00:11:22.920 | And in this train.py, you effectively have all the training code you need abstracted.
00:11:28.040 | It's about 500 lines for walking.
00:11:32.600 | Yeah.
00:11:32.920 | And then, basically, you can run this on like, you know, using run-power your local GPU.
00:11:37.480 | It's MJX.
00:11:38.200 | So, it's like, you know, it's accelerated, accelerated compute.
00:11:42.520 | And training and walking policy roughly takes one hour to two hours.
00:11:46.840 | And, yeah.
00:11:47.480 | So, we're just going to run through like millions of different, not examples.
00:11:51.000 | Yeah.
00:11:52.360 | Iterations of the robot performing tasks you want.
00:11:54.600 | And you can tune the reward functions and et cetera.
00:11:57.560 | And you can see the loss and reward functions in our observability.
00:12:02.200 | Basically, tensor board.
00:12:03.320 | Yeah.
00:12:05.480 | And afterwards, when the robots finish training, you can easily evaluate it in KOSM.
00:12:11.080 | So, all you have to do is like, k-inverse-sim, like, you run the policy.
00:12:17.720 | Yeah.
00:12:18.040 | And in simulation, and you see the robot.
00:12:20.200 | If it's walking, if you can see if it's doing the thing you want it to be doing.
00:12:24.280 | For example, like walking, standing, picking up objects.
00:12:28.440 | And if that's really good in simulation, then you can just easily change the IP address.
00:12:32.920 | And then, you have sim to row deployment.
00:12:35.080 | Yeah.
00:12:38.760 | And it's very cool.
00:12:40.840 | Like, you can basically get a robot to work in like, one-tenth of the time of like,
00:12:45.800 | what you would take to set up most training libraries right now.
00:12:49.480 | And also, like, we're a team of 15 people.
00:12:53.480 | We do everything from hardware to software to ML.
00:12:56.040 | So, how we're able to do this is actually by working with our open source community.
00:13:01.080 | Currently, we have about like 5,000 active Discord members in a few servers.
00:13:07.000 | We have a lot of public open bounties that people are tackling.
00:13:10.760 | And because of our software's MIT lessons, yeah, a lot of people are coming to help us.
00:13:16.440 | And we also run hackathons almost on a bi-monthly basis that a lot of people come to participate.
00:13:21.560 | Yeah.
00:13:24.120 | So, we're also hiring electrical firmware and ML engineers.
00:13:27.800 | So, if you're interested, feel free to ask me.
00:13:31.800 | And then, also, go on the kscout.dev/joint website.
00:13:35.160 | Yeah, we're trying to hire a lot more cracked people to join us.
00:13:42.040 | So, we're launching the robots and software stack in about two, three weeks.
00:13:45.400 | If you're interested, follow on the website.
00:13:48.200 | I'll be happy to answer any questions.
00:13:57.960 | Yeah, sounds good.
00:13:58.680 | Go ahead.
00:14:00.040 | Yes.
00:14:00.280 | Where's the power?
00:14:03.720 | The battery?
00:14:04.120 | Is it battery packed?
00:14:05.640 | Yeah, it's battery packed.
00:14:07.080 | Yeah, it just has a battery.
00:14:08.600 | I don't know if I can show you in the picture.
00:14:10.920 | Yeah, it's behind this, basically.
00:14:13.080 | You can just slot in and it clicks in.
00:14:16.040 | Yeah.
00:14:16.840 | Yeah, the weight of the battery versus longevity.
00:14:30.680 | What do you mean?
00:14:31.080 | Yeah, longevity on the chart.
00:14:32.760 | Oh, like how long is?
00:14:33.720 | Yeah.
00:14:34.440 | Yeah, walking so far, our test is about two hours.
00:14:38.280 | But you can pass through.
00:14:39.800 | So, you can power through the wall plug.
00:14:41.800 | So, it's, yeah, you can just keep letting it charge and also run at the same time.
00:14:47.400 | Yeah.
00:14:47.640 | Yes.
00:14:49.880 | Well, black jacket, yeah.
00:14:53.000 | What are the use cases that you guys are imagining to start off with?
00:14:58.920 | Is this going to be more for like commercial, like, you know, factory kind of use cases?
00:15:02.680 | Or do you envision this more in the home, like helping out?
00:15:05.640 | Yeah, so basically, our bet, so a lot of companies, like especially in the US, are betting on like B2B.
00:15:12.760 | So, like Figure, for example, are selling to factories.
00:15:15.000 | Sims, Tesla itself is the customer.
00:15:17.560 | For us, our really bet is becoming the first US consumer robotics company.
00:15:21.960 | Like, a robot, a humanoid robotics company.
00:15:24.920 | Yeah, so we're really selling it to anyone that's interested in developing robotics.
00:15:28.920 | So, a lot of our current customers, we accidentally launched our robots.
00:15:33.320 | Like, people accidentally started buying our robots through our Shopify page.
00:15:37.080 | That was a complete mistake.
00:15:39.000 | But a lot of people that bought were just people genuinely interested in using for household tasks,
00:15:44.680 | like programming for different research.
00:15:47.000 | And also, there are a lot of companies also interested in working with us to make B2B businesses.
00:15:52.840 | Like, for example, the food demo that we just saw.
00:15:56.040 | Yeah.
00:15:57.640 | Sorry, if I can ask a follow-up question.
00:15:59.080 | Like, what household chores do you think it would be well suited for?
00:16:03.240 | Like, unloading the dishwasher, for example?
00:16:05.560 | Yeah, yeah.
00:16:06.200 | I mean, right now, we don't have any, I don't think anyone really has a fully working VLA model yet.
00:16:11.720 | So, right now, it's pretty limited to teleoperating system, sorry, teleoperation.
00:16:15.800 | So, yeah.
00:16:17.000 | But soon, we hope to be able to have, like, this navigation VLA stack for you to do, like,
00:16:22.040 | you know, folding clothes or, like, doing dishwashing as the model capabilities improve.
00:16:26.040 | Yes.
00:16:28.040 | White shirt?
00:16:29.800 | Oh, I mean, green shirt.
00:16:33.160 | Yeah, you can go first.
00:16:34.040 | Can I ask you?
00:16:35.960 | Yeah, of course.
00:16:36.680 | Okay, yeah.
00:16:37.720 | So, you kind of alluded to complexity of ROS2 in terms of the setup.
00:16:43.320 | Uh-huh.
00:16:44.280 | I'm wondering if there are other, like, benefits and trade-offs that you considered
00:16:48.440 | for foregoing something like ROS and ROS2 in that ecosystem?
00:16:51.720 | Oh, yeah.
00:16:52.120 | Like, why not use ROS, basically?
00:16:53.720 | Yeah.
00:16:53.720 | Yeah.
00:16:54.280 | Well, there are a lot of reasons.
00:16:55.880 | Our robot is mostly programmed.
00:16:57.320 | So, ROS is really good because, like, the nodes and stuff, right?
00:17:00.120 | So, like, the async, like, the communication.
00:17:02.440 | But for our robot, we really don't have that many sensors.
00:17:05.720 | And we really want to do this, like, model-based, like, policy-based robot.
00:17:09.720 | So, we don't have many complicated sensors that we need to, like, async-communicate at all times.
00:17:14.600 | The other part is, like, where I'm pretty opinionated.
00:17:19.080 | I used to ROS1 and ROS2, Foxy and Noetic.
00:17:22.520 | I've just had a pretty bad experience using it, having to set up Ubuntu, you know.
00:17:27.640 | Like, I just want a robot that I can just buy, open the box, it stands or walks,
00:17:33.240 | and then I can just start programming it using my computer.
00:17:36.120 | Yeah.
00:17:36.360 | What kind of AI accelerator is on each of the robots?
00:17:46.360 | Yeah.
00:17:46.840 | So, basically, for the KBOT, it's going to be JSON, Nano, and AGX.
00:17:51.720 | Yeah.
00:17:53.720 | So, yeah, there are different compute options you'll be able to choose when we launch.
00:18:02.280 | Yes, yeah, go ahead.
00:18:03.240 | How do you develop for this?
00:18:05.160 | So, you can just put either VR headset.
00:18:07.240 | So, there are a few different methods.
00:18:09.400 | So, the preferred option for a lot of people is VR headset.
00:18:11.960 | We have this, like, we also turn, like, a pseudo-IK.
00:18:15.880 | So, basically, it's, like, going to position using, like, an RL model,
00:18:21.080 | instead of just, like, calculating an IK.
00:18:22.600 | But, it works pretty well with our VR setup.
00:18:25.000 | So, you can move the hand gestures, you can click button to open and close gripper,
00:18:28.840 | and just, yeah, move your arm and stuff.
00:18:31.000 | Yeah.
00:18:31.720 | Can you still operate the small ones here?
00:18:33.560 | Yes.
00:18:34.120 | Yeah.
00:18:34.440 | You'll be able to.
00:18:35.320 | It runs the exact same software stack.
00:18:37.000 | Yeah.
00:18:37.800 | Last question.
00:18:38.360 | How does it compare to the Tesla?
00:18:40.920 | The Tesla humanoids?
00:18:43.480 | In terms of, like, mechanical powerness, like, you know, the Tesla's way more powerful.
00:18:49.160 | It has, like, linear actuators and et cetera.
00:18:51.480 | But, in terms of, like, actual use cases, I don't think it's really that different.
00:18:55.320 | Yeah.
00:18:56.040 | I mean, Tesla is actually built for, like, a factory type of use cases.
00:19:01.000 | But, in terms of, like, you want to, for people to actually buy and use this robot, it's not very different.
00:19:05.400 | I think the last time I heard Tesla Optimus is about 60k, at least.
00:19:12.200 | We can ask some Tesla engineers.
00:19:14.920 | But, our robot's $9,000 before mass production.