codekansas

I've been thinking for a while that some part of the ebullience towards AI among the American decision-maker class is that it is a good way to stick our heads in the sand and pretend like super-intelligent AI will make up for not being able to build competitive cars and drones. It has the feeling of an easy-to-digest explanation but I'm worried that it's probably wrong.

But they're bootstrapped and using their own money, not defrauding investors

I suspect that these kinds of things will happen with or without my involvement. Assuming that they do, I would rather that a good open source option exists

Oh I see. Yes, we're currently exploring a few different five finger hand options - we will choose whichever option provides the best value. I actually quite like the Inspire hands and we might be able to get a volume discount

Really appreciate it :)

I think we will just focus on making really great, low-cost hardware and a nice SDK, and let other people experiment with different approaches on the intelligence layer.

Well, I believe the competition up until now has mostly been on hardware, but moving forward it will mostly be in software. I don't think we will be outcompeted by Unitree on software. And I hope to capitalize on our engagement with the open source community in a way that Unitree has not.

We try to sell it for a fair price while still making money. Actually I think there will be comparably priced humanoids coming out from other companies soon

I really do think that building through the open source community is the best way to compete with the big players, even without having a lot of capital. Of course, it doesn't mean we can't execute well, but I do think it's a good way to make a lot of progress without spending a lot of money.

I think it's a very hard problem, but I would like to align our company's incentives towards solving it

Yeah, the basic robot is just a robot, albeit with open-source software and hardware

For compute - we're exploring a few boards right now, but the base model will be something from Amlogic and the higher end model will be something from Nvidia

My overall plan is basic joystick control -> VLA with RL -> self-supervised embodied representation -> end-to-end RL -> end-to-end control. I suspect there will be some very good multi modal models coming out in the next few years which we might use as base models, although more likely, we will adapt their techniques to work on data from our own robot.

I agree that the hardware is easier than the software - I am a software guy, personally, but I felt that it was important to do the hardware first so at least we can have a baseline product which we can offer to people. I would personally like to work on this software problem (or rather, build a company to work on this problem), and this seems like the right way to go about funding working on this problem.

This is really interesting to read about. To be honest, I know very little about this space, but it's something a few people have approached me about tackling.

I do think that this is a great application of a general purpose robot. I'm not sure what the technical timeline will be, but it would certainly be cool for my parents to have such a robot when they are elderly.

We're exploring some options - maybe a rental option in the future. Would like to make it accessible to everyone

I have a pretty bad mental model of how most VCs think, but I think good VCs will fund smart people who demonstrate extreme conviction, regardless of how they initially size the market. The opportunity cost for me doing K-Scale is making quite a bit of money at Tesla or Meta, so assuming I am not acting irrationally, either I have extreme conviction or I am a masochist. In my experience, VCs are pretty bad at telling the difference.

Thanks!

We're all pretty cross-stack - there are some hardware people and some software people, but the product is quite integrated. Personally, my time has been mostly focused on the RL stack recently, and after there are more robots in the wild, I suspect my time will transition to working on building this data feedback loop.

I try to answer questions pretty actively on our Discord so happy to chat there about whatever you like

Which link? I can fix

Yes, we are aware of this customer segment. I don't think they have thoroughly considered the implications of what high torque actuators can do to the human body

Yeah I think humanoids are a pretty fraught area. There's definitely been some backlash but overall I have really appreciated the responses we've got from people. Like, I just want people to care about what we're doing - if positive that's great, if negative then we can learn how to do better. As long as people have an opinion one way or the other, I'm happy :)

This is a much better answer than mine

Yeah of course. That's kind of the whole point - I don't think you can really trust a humanoid robot in your house around your family if it is not clear what it's running. Basically, for myself, I would not want to buy a closed sourced humanoid, and I view myself as relatively representative of the early adopter mentality. So personally I think this is the right way to build a great product.

I basically believe that in a world where humanoid robots are actually useful, we will not have any trouble monetizing. Probably we will verticalize manufacturing at some point in the future. I think the bigger risks for our business model are not from people copying us or something, but from not making progress fast enough.

Yeah, this is exactly the kind of use case we intend to support. Basically, we want our robot to be the best mass-produced robot for this kind of experimentation.

Yep I definitely agree. Our team is pretty small and bandwidth limited

I basically think our goal is to solve all the boring stuff and make it work reliably, so that other can people can try out the cool ML stuff more easily

Eh. I think we got a bit nerd-sniped on some things and we ended up trying to build most of our stack ourselves. The control loop is just a pretty simple Rust loop for running ML models. ROS is kind of annoying to work with and the control loop is pretty simple so we just wrote it ourselves.

For two legged - I think it will eventually be quite a bit cheaper, it's mostly a software problem to get them to be stable. RL based control has gotten much, much better. For example, I was talking to Trevor Blackwell a few weeks ago, and he was saying the IMU on the original Anybots robot was over $2k, whereas if you throw a noisy IMU into sim, you can get away with something basically from a cellphone. So yea, personally I'm a big believer in needing to solve the robotics intelligence problem, and once you solve that, humanoids will make the most sense as a form factor.

Actually yea, the benefit of our parallel gripper is that we get some proprioceptive feedback which we can't get from the current 5 finger hand. I'm not sure how important this will be long term - I think vision can eventually mostly compensate if the ML models are good enough

Tao Motors reached out to us about a month ago after our soft launch because they just bought a big factory in Texas to manufacture golf karts, and the supply chain for golf karts and humanoids is actually pretty similar. Their subsidiary invested some money in our company recently and we are partnering with them to scale manufacturing. I discussed this in our launch video.

So, most humanoids you see that are relatively cost-effective are just "humanoid" in that they look like humans, but there are significant mechanical differences between them and real humans. It's almost always driven by the cost of manufacturing different components. A good example is the lead screw you see in the knee and ankle on Optimus - while it is more human-centric to have tendon-like actuation, it drives the price up quite a bit. Put differently, humans have a lot of specialization which is not great if you want to mass manufacture something.

For walking, the most important thing is that the robot can be simulated well, so in our case, we tried to model the foot contact with the ground in simulation quite accurately. In fact, we found that force sensors in the foot probably help but they're not necessary in simulation, and we wanted to shave off anything that wasn't necessary. I am not sure how close we will get to human mobility - it's definitely a learning process - but you can get much further in simulation than you'd expect.

I think it's tricky to manage an open source community effectively while still being opinionated on the product. I don't want to get too sucked into coordinating adhoc contributors - we do have a strong core team of people, and we all live together, which is pretty important. But yea, still figuring out what the right dynamic here will look like. Thanks for the suggestions!

To be honest, I don't know that there are many applications today which humanoids are the best form factor to solve. I would view it more as a form factor that is likely to get much more useful over the next few years, and having the hardware in your home lets you try out new techniques as they come online.

Personally, I think the first real use cases will mostly be entertainment. Humanoids have a high "coolness" factor. Also, I think there's a long tail of random problems which you don't want to buy a new robot to solve, but which, if you have a robot lying around that isn't perfect but is "good enough", might be possible to solve imperfectly. For example, I just had a newborn baby, and I was thinking it would be nice if I had a static arm that could hold his bottle for me. There's a lot of tail end problems like that in your day to day life. But I think the really interesting capabilities will come once there's very good end-to-end models running on-device.

Longer legs means more inertia on the actuators, which means bigger actuators + bigger battery, which means heavier robot, which means...