All right. Um, welcome to the afternoon session of Alignment 2025 and uh, I hope

you guys had a great lunch and welcome to the robotics panel. I'm your host

Jing Wu and I just introduced um, you know, our three amazing guests over

here. The fun story I said I was going to tell is, you know, um Anand was here last year and after he came here, this year his company actually had a $800 million $800 million fund uh used for robotic startup financing.

So I guess my point is I really hope our event can be a lucky charm.

So we saw a really amazing demo, man.

Um I think um you know part part of what Chen introduced is we have different ideas in this panel, right?

So Anna, let's actually start with you.

Um uh you're going to be our like X factor here because you know it is an amazing demo but honestly I can't remember when's the last time that a robot demo does not have folding laundry or picking fruits.

So, I guess in an environment so hyped up about humanoid robots, how can you still stay so poised and motivated to work on cleaning robots?

>> That's because I make money.

>> Okay, that's a punch line.

>> I have large enterprises throwing a crap out of money to solve real problems and you can, in my opinion, you fall into two camps.

You can either get revenue from customers by solving problems or raise millions from VCs by chasing hype.

>> Yeah, that's cool.

So if if real application or task specific robots can make money. So I'm not sure if we mean human or robots cannot make money down the road, right?

And that's that's a sharp perspective.

I guess Evan, you just did the demo.

Um you know, how's how is it like working with Elon at at you know Tesla Optimus because you your background you you were from uh the Optimus team.

Um how's it like working with him and how did that experience influenced you to do your own startup as well?

>> Yeah, absolutely.

I think Elon is definitely a great and one-of-a-kind entrepreneur and he's also has this uh deep techdriven uh mindset and he is a true believer of first principle.

So we uh from dayto day uh basically as Optimus since it's his uh main focus these days

we had a kind of weekly meeting with him to go through the updates and his directions overall.

He always tell us to always trying to get the foam human form factor even though it's uh it's going to be a long pass and challenging experience to get there but he truly believes like once uh the technology there uh eventually human uh we will get there and uh also the market will be enormous and yeah that's and we we sometimes even it was him at like 11 p.m.

So, so he come to the office to to review the robot at 11 p.m. Everybody's stay.

So, so he's a really hardcore startup founder, I would say.

>> Wow, that's amazing.

Does that mean your company also work until 11 p.m.?

>> We are building humanoid, but we are human.

>> Ah, that's that's perfect answer.

Yeah, if you guys are hiring, then I'm sure the folks will be happier that to know you're not, you know, 996 or whatever.

Um, but yeah, coming to you Yan, I know um well, speaking of hype, right? And um I I read somewhere that Openmind's actually your company Openmind um they put Asimov's law and the robot identity actually on blockchain.

So if that if those words coming together is not hype, I don't know what is.

Um but I actually I I know you saw the real deal though because um you were already a a professor at Stanford.

So I was wondering what do you see in humanoids now that actually make you so excited to come out and start your own company when you're already have so much prestige and you know resource at Stanford already.

>> Uh well uh there's a lot to unpack in your question.

Um, but uh what's exciting to me as someone who uh loves robots is that uh the long-standing problem in robotics, which is to have a clear line of sight to quote unquote um universally capable uh humanoids, uh is now there.

And that's because most large language models uh speak fluent robotics.

So this old perspective on how to build software for robots which is either sort of deterministic task focused or endto-end AI now we have uh another set of tools that allow us to rapidly repurpose um robots for a diverse array of tasks in hospitals, homes, schools, workplaces.

And that's the really exciting part is that large language models, many of them speak fluent robotics already.

>> What do you mean by speak fluent robotics?

Do you mean like they they have good connection with the code API or you know do you mean humans can use language to command the robots better?

>> Well, one way to think about this is that large language models are very good at completing stories.

If I give a large language model the first chapter of a book, I can ask it to write the next 11 chapters.

And if I represent the world as a uh contiguous block of text, I can then ask a large language model to continue the story.

If I then connect the large language model to uh a physical shell that allows it to move, explore, interact, bark, ask, show emotion. um then all of a sudden I have this object in front of me that is uh much more interactive

um and uh much more engaging and can also learn in ways that really hasn't been true before for robotics.

>> I see that's a good point.

So the interaction side of things um an is there anything any quick comeback or anything quick to add um after you hear from two folks that are you know b slightly bullish on on humanoids maybe do we need like both task specific robots and humanoids in the long term or

>> depends on what part of the sector or industry you're going after is what I would argue.

if you're going after B2B enterprises where people have enough money task specific is far better because just that single task alone will take more than 24 hours for a single robot to complete

>> and then I don't need a general purpose robot that's overbuilt.

An analogy I like to make is if Henry Ford was asked, okay, how are you going to design a far better transportation system? He's not going to take a horse or a human being and make that into a system.

No, he's going to take a car.

Cars are far better at transporting goods from A to B.

My vacuuming robot will far out compete any vacuuming humanoid or humanoid trying to vacuum with a Dyson.

So this building alone, if I had to vacuum this entire space, I'd need about four or five different vacuuming robots to be able to take care of that.

It's far cheaper, far more reliable, far easier to be able to get that done than to have a single humanoid trying to do or a team of humanoids trying to do the same thing.

So it all boils down to what are we going after?

And in enterprises, tasks are meaningful enough, large enough that an application specific gets you to your ROI far more reliably.

>> Yeah, that's a perfect point and I I believe you because some of Anna's clients actually include some really big names that we know, you know, um big big tech companies like, you know, Apple or these companies in the in the valley are your clients.

So definitely um so it seems like the key is how do we roll out robots to its uh proper applications.

What are the roadblocks to um the mass adoption of robots, right?

So I guess on the mechanical side coming to you Evan um kind of you know I your your hand is really amazing because we might not know it's it only cost $300 and it squeezes in the features and the degrees of freedoms of hands that usually cost thousands of dollars.

So I wonder how how you guys are able to make it.

Are there any recent breakthrough and what what still makes it so hard for you guys to you know um do some of these complicated tasks?

>> Sure. Absolutely.

So I think coming to a robotic hand design especially human form factor it's not only about uh you know like we just need to uh get this joint to move this many degree of freedom.

It's more about understanding the end result the functionality of a human hand why it's such a form factor and how people are interacting with object around you.

So that's why we are uh sticking with this uh cable tendon driven form uh architecture and I know there are some other like different architecture like linkages or direct actuator drives those are uh like you see them move smoothly and nicely but they are not built to really uh like work like a human hand.

So that that's kind of very important that's why we picked this cabled driven architecture.

Uh yeah that's uh my two cents from uh from mechanical design perspective.

Yeah. Yeah that's great.

So I guess then from the software perspective because Evan just said right we have different software based on different type of hardware.

So that's also kind of the uh what companies usually do um if figure or if um you know um Nvidia their robots is built in a specific way they're going to have a specific model for that.

But Yan, you believe in open- source and modular software.

You believe in different types of hardware might be able to work with the same open source software that you have.

So I guess why do you believe u the robot operating system needs to be open source and yeah how how does that help solve some of the bottlenecks in software?

Okay, well um again there's loads going on uh in your question.

Uh there's uh why is open source useful?

And of course the internet runs on open source software and most phones on earth run on open source software.

It's called Android.

So there's certainly um a lot of examples for how open source software powers really important things.

Um the other thing you alluded to is um the uh optimal architecture to use for thinking machines.

And uh as we've already heard in robotics, it's extremely important to be extremely specific.

>> When people throw around things like robotics or humanoids, um it is vital to be specific about the task you're trying to accomplish.

Um, for example, um, if you're building a humanoid sex robot, obviously you're trying to accomplish a very different set of, uh, like, uh, it's a very different business case and form factor compared to a vacuum cleaner compared to something BMW would like to use to put Torx bolts into their engine control units.

If you think about the global economy, uh something like a third of that is based on uh manual labor like people doing things.

It's something like 30 trillion a year is based on humans doing stuff in the physical world and something like a third of that third involves hands.

So obviously if you're going after something like iPhone assembly then uh exquisite hands are critical.

So um then with respect to your question um what is the right architecture for humanoid robots?

There's no simple answer.

Again it depends on what you're trying to do.

Um if you're physically manipulating things that is requires exquisite techn technology like we saw earlier today uh with hands but it also requires amazing compute.

Yeah, the difficulty of making sense of your physical world, deciding what you should do with the hand, and then being able to deal with all those degrees of freedom to accomplish a task is a uh big compute problem.

with respect to what a lot of companies are doing right now which is exploring the possibility of modular Lego block like architectures for humanoid including um intrinsic including OpenAI's robotics including Nvidia Foxcon humanoid including everyone else who's currently working on software for humanoids the central idea is that it's much easier to debug smaller pieces of software rather than uh like fine-tuning some end to-end AI or something like that.

So it's not only us that is thinking about modular architectures, it's many other people in in the industry as well.

>> Yeah. Yeah, that's a great point.

I guess Evan coming back to you very quickly because you working on hands before um you know right now on robotic hand we probably saw the demo succeeded one time today but yesterday uh you know we there's like a number of tries and we were talking about how difficult it is with the you know uh tactile sensors or there there's um maybe a lack of force feedback on the hand.

So what are your thoughts on some of these uh hardware working with the you know smaller uh modular software like professor Ya mentioned?

>> Okay. Okay.

So uh I think your question is about uh like uh uh some hardware sensing and

>> yeah like software working with different type of hardware like do you guys write specific software for your hardware or can do you see it being able to work with other types of software in the future?

>> Absolutely.

So with the hand we showed today, it's a total pure open source hand including the hardware and also software.

We are going to rely on industrial and academic um people and even hobbyists trying to uh contribute to this whole thing and to uh create this ecosystem around the hardware to make it better and better.

That's why we just marketed at $314.

So like it's it's a really easy to enter this uh dextrous manipulation world without like breaking the bank.

>> Yeah. Yeah.

That's a good point.

Um I don't I guess come to you.

Um you've seen robot deploying in the real world.

What do you feel like is the bottleneck um to the mass adoption of robots?

Is it on the hardware side or on the software side or maybe it's something else?

So between us and our partners we have about 35,000 robots deployed in the real world.

Clients include Walmart, Amazon, Apple, Nvidia, Meta, Hilton, Marriott, MGM, etc.

>> The challenges that BC the biggest challenge today that we are actually overcoming is not building the robot.

It's not the software.

It's not even the hardware.

It's the fact that robots break down all the bloody time.

Question for you, Evan.

How often do your robotic arms break down?

Yeah, it's uh so so yeah I would imagine once we really deployed it that's a quite a big issue at this moment.

>> Yeah. So in the 48 states we are deployed in today, pick any state, I will tell you a city right now I have a broken robot

>> and that is my biggest headache

>> is uptime being able to make sure my robots are operating to that end like we are hiring thousands of veterans to go out and train them and be a robot technician.

>> But that is the real world challenges that need to be overcome.

Along with that there's the entire finance aspect to this like how do you actually get the capital to be able to do all of these systems?

Robots aren't cheap.

Hardware is not cheap.

>> Yeah.

>> And then a lot of the tasks people go after such as laundry folding is this is minimum wage tasks.

Most parts of the country we are still at $12 $10 per hour.

So how do you make the math work in that front?

>> Yeah absolutely.

So you talk about business model and reliability and that actually are the two more talking points that we're going to get into.

I guess in terms of business model, Evan, um Elon last year said the Optimus robot will cost only around 20 to $30,000.

Do you feel like that's feasible?

And if it is, how can you know startups like you guys to compete with companies like Tesla because they have so much more manufacturing capacity?

>> Yeah, totally.

Uh first of all with the price point Elon brought up a 20 $30,000 I I'm truly believing that like uh once anything get to mass production we are basically uh optimizing all this process and so the cost down to basically bomb cost and pure material cost and talking about the cars right cars have such a high mass of material around it but it's still maybe around 40 50,000

But so the humanoid robot eventually I totally believe we will get down to 20 $30,000.

However, uh so so talking to how we can compete in this game versus Elon or or Tesla uh I believe as Aneno was saying like uh in the future uh I also believe there will be many different form factors of robots.

Uh so Elon is trying to build the Optimus which is one SKU right like one human form factor but a human is not designed to u be the most efficient form factor to do all kind of task.

So um but the hand I believe will have more uh compatibility to on different tasks.

So it will be different platform but integrate uh all kind of hand shape and as a hand uh focused company will also developing you know different skills kind of hand even the human hand we build differently like uh some people with very elegant hand can do uh like music instrument and some hands are more like uh like strong and like labor driven hands.

So, so where our robot hand will be designed.

>> Yeah, maybe one day we'll see, you know, the chairs being carried on by robot hands and with the robot themselves.

For sure.

Um, Jan, I know you mentioned Android and robotics.

I guess part of the reason why Android works is because it also integrates closely with the rest of Google's ecosystem, right?

Like it drives people to Google's ads revenue and all of that.

But for a startup who does not yet have that type of ecosystem, how do you see open-source business models start to work out in robotics?

To uh quickly comment on what we heard earlier, uh there's some really excellent points uh we're surfacing.

Um, one of them is the reliability and I totally agree that's probably the main thing right now is just um we expect when we go outside we expect our Tesla to just work and uh most robotics deployments right now um except for things like in the Gigafactory or in car factories uh have nowhere near the robustness or reliability uh that one would like.

Um, the other thing we're seeing as a practical barrier to deploying, uh, humanoids and quadriped into American households, uh, is as simple as liability insurance.

Um, try to go to an insurance company and say, "Hey man, I'm deploying humanoids into US households on October 15 for healthcare, math, education, and security.

I need liability insurance."

And they'll laugh.

So the good news is that Munich Ray as of last week is offering AI shore which is liability insurance for humanoids.

And the point of that insurance is that if the humanoid fails to perform a task uh Munich Ray will cover liability arising uh from that.

And so totally agree with um Anad for certain specific use cases the technology the hardware is basically good enough uh the software is basically good enough.

Um, it's really things like business models.

Is this leasing?

Is this a subscription service for add-on apps for your humanoid?

Is this something that old people's homes purchase for their patients?

So, who's buying?

How much are they paying for it?

How are they paying for it?

And how are you dealing with things like uh reliability and uh insurance?

There's one new task for humans that blows my mind and this is based on deploying humanoid into memory care facilities.

So that's a situation of where the humanoid does not have hands because it's not necessary.

The point of the humanoid is to engage with a human and make them laugh and cry and listen for people who've been in memory care facilities uh for years and have almost no interactions with other humans.

And humans in that situation get so attached to the humanoid that the nurses in the evening have to wipe lipstick off of the head of the humanoid.

And uh so this is part of the general like maintenance of robots deployed into human settings.

>> Yeah, I see.

Well, I was looking for different opinions, but I got I got a lot of agreements with, you know, between you and Anna.

So, I guess that's a that's a great point.

Anna, you were smiling a lot before um when when Jan talk about like insurance and all the all these problems.

Do you have anything quick to add there?

>> Yeah, that is what I'd say is the difference between the real world and then like the academia or the startup world.

In the real world, insurance is a big thing.

COIs, being able to comply legal uh good luck trying to get a large language model approved for any of these large companies.

Good luck trying to get Chinese robots like a unitary inside where you know there are so many back doors where they can the privacy is a big aspect.

Data security is a massive aspect.

like how do you get them deployed in large scale industrial sectors?

How do you get offices?

Uh we clean Jamie Diamond's personal office with a robot.

I am paranoid cuz I was like holy that's a huge security risk.

We are in multiple defense sectors.

We have robots running around there at SpaceX and at military bases.

So how do you get through the security clearance?

How do you get through insurance?

How do you get through reliability?

How do you get through making sure that when the robot breaks down, someone can show up and fix it?

like that is the real world that needs to be solved far more importantly and people are spending a lot more time trying to figure those portions out.

>> Yeah. Yeah.

Absolutely.

And we're going to we're going to cover reliability very quickly but um first on more about your $800 million fund.

I guess one thing is how did you even discover that as one of your um business model?

Um where did you find that need and why can't just robotic startup find customers and funding on their own?

It's a great question.

Uh, longer story short is getting access.

So, we spoke with 30 plus CEOs of robotic companies who all raised a series A and could never get to a series B.

And this was generally like I'd say the wave two of robotic companies.

This was the Willow Garage style, the Savio Robotics, etc.

Some of the most legendary robotic companies there are.

And one of the biggest reasons they fail is getting access to capital.

Capital is expensive.

Typically, there are two ways of getting funding.

You can go to debt, your bank, Silicon Valley Bank who not incredibly high rates because you are a startup, you're not profitable.

Or you can go to VCs and VCs will take a portion of equity, but that's gated capital.

If you don't hit certain benchmarks, they're not going to give you more money.

Now, robots are expensive.

They're not cheap.

They're not easy to build.

They require a lot of R&D.

And most of my customers don't want to pay upfront for $50,000 for a robot.

what they would much rather do is $1,000 a month

>> because that's what they're used to paying a human being to doing that particular task.

So how do you bridge that particular bridge when if you get a order for 500 robots and that robot cost you say 20,000 you require $10 million worth of capital to be able to even go out and buy or procure or build so many robots?

Who is going to give you that $10 million?

The VCs?

No, that's too expensive.

Equity dollars don't exist for that.

and banks they will eat up your entire margin.

So we created this fund with a consortium of 18 different banks to be able to actually underwrite the contract, underwrite the end customer and do it at what 3 4% compared to what Silicon Valley Bank gives it at 18%.

>> Yeah, that's a good point.

It sounds like a lot of effort trying to go uh go in to enable um us deploying more robots in the real world without worrying too much about the underlying risk because that's always a work in progress.

So speaking of reliability, uh I don't you've covered it a lot uh on the reliability side.

I want to um ask Evan as well.

So um you know h how far away are we from reliable robot hands?

And also um one thing extends from reliability is if the robot is reliable, it's safe, we can start to trust them.

So, Professor Ronnie Brooks actually said in one of his essays earlier this week that if the humanoid hands have tactile feedback, it can really help foster a better human robot interaction.

So, just largely how does robot hand all play into all this?

I I would say from my uh two cents maybe in the next uh uh two or three years we're going to see a lot of maturity on the hardware side especially on the hand but for the general whole body especially the bipad right that that still uh takes maybe a little longer because to get to that level of maturity need a whole system improvement but just by hand itself I totally believe when uh with the whole resource from the whole society focusing on getting this better and better.

We believe is we'll get there in the next two three years.

>> Wow, that's an optimistic uh look.

And I guess Yan um in terms of reliability, like I said, is is is trust uh is interacting with the robots.

You're actually one of the only maybe the only person I know that lives with robots.

I wonder um how are Iris and Fronti doing?

And um yeah, how do you ensure safety and um trust when you actually interact with these robots that actually live in your home?

Um well uh I encourage all the engineers and investors and bankers and scientists in this audience to get a robot and try to live with it.

Uh because that will immediately expose a long list of things that are irrelevant in the lab but critical for deploying humanoid or quadriped into households.

A great example would be cooling.

Um, a lot of the humanoids right now have these big fans going

>> Yeah. Yeah.

24 hours.

And if you've ever tried to sleep next to a humanoid with a loud cooling fan, it's literally impossible.

That's irrelevant if uh you're a humanoid in a lab because no one cares about sound levels.

But the other thing is when the humanoid walks around your house, it's very loud on wooden floors.

It goes clunk clunk clunk.

And so those are little things that uh become only apparent when you try to live with a humanoid.

And with respect to trust, um the whole notion of building open- source software for humanoid um for me personally was a like no-brainer.

Um I have kids.

Um I want my kids to be able to look into the brain of thinking machines around them and say, "Aha, um I can kind of see what's happening here.

I can see how the data are moving.

I can see bugs.

I can contribute, add, fix.

And in retrospect, um, focusing on open source software for humanoid, um, was not only meaningful for me as a parent, it also makes good business sense for reasons we heard earlier.

Imagine you're a quote unquote international manufacturer of humanoids and you're trying to sell into US markets.

Like, good luck.

Um imagine uh like you know you want to sell your humanoid into a US uh police department, fire station, elementary school, hospital and the first question you get is oh you know why should we trust the software?

Where are the data going somewhere I don't want it to go?

And a good solution for that is that if you build good hardware, why don't you use open-source software brewed with love in San Francisco and have the data stay in the US on **HIPAA**（Health Insurance Portability and Accountability Act: 美国《健康保险流通与责任法案》，旨在保护患者医疗信息的隐私和安全） compliant servers?

And so it turned out that building open-source software for many many different verticals is not only good for like all the parents in this room who care about where the technology is going but make very good sense nuts and bolts for **收益分成协议**（revshare deals: Revenue sharing deals，指合作方之间按照约定比例分享收入的协议） with Chinese robotics companies.

>> Well um we got like 30 seconds left.

Anything quick to add on reliability or or uh the safety trust side of things?

>> I have one question for Evan actually.

Evan, where do you see the need for five fingers versus three?

Like what task truly requires five fingers that three fingers cannot do?

>> Yeah, I kind of asked him the same question yesterday actually.

>> Yeah. Yeah, that's actually a great question.

I I think uh when you talking about basically pick and place task, right, three finger is totally enough for like tripod support.

However, once you start talking about uh like doing tool operation like using a vacuum, right?

You need at least four finger, I would say four finger to to to to hold your tool first and then use your index finger to pull the trigger.

If if you don't have this extra redundancy as your ring finger only three, you're only picking up your uh vacuum with two finger now and then one index finger for pulling the trigger then that's not stable state.

I I I would argue maybe fifth fifth finger is not that important on the pinky, but four finger is definitely a need.

>> Yeah, this cool.

I I love how this turning into a biology and evolution theory class real quick.

But um yeah today we covered uh robotics uh progress and challenges from the trinity the impossible trinity which is the technology and the cost and also the reliability side.

So yeah I really appreciate our guest here and let's give it up for our panelists.

Thank you.

>> I know I have the answer for you.

So if you have a robot girlfriend you need the ring finger

>> right?

How do you propose to the girlfriend without ring finger?

>> That's right.

That's right.