the future where you have humanoid at home folding your laundry is a lot closer than you think and the price will also be a lot lower than most people imagine we can manufacture this at a cost of relatively affordable car at 1X we're focused on building human form factors to do many tasks in the home and we think that this will lead to human level [Music] intelligence for the first episode of year 2 of s 3 we're getting an exclusive look at 1X Technology's new robot Neo 1X has been building robots in Norway since 2014
developing their own tendon Drive technology as well as their first robotic system Eve a humanoid-like robot with a wheel system for ground movement and parallel grippers for hands but now onx believes they're not only ready to build a humanoid robot but manufacture them and deliver them to paying customers for use in their homes as soon soon as next year I drove from our studio in San Francisco south to Sunnyvale to meet Neo for the first time and asked their team more about how they plan to bring humanoid robots to the home my name's burnt I'm
the CEO and founder onx and we make humanoid robots for the home it's a cliche but I knew I was going to do this since I was a kid kind of kid that had like his cabinet full of stuff that I picked apart putting it back together for some reason one day while like looking at an excavator doing work and like okay I'm going to make you or robots that's what I'm going to do in life I think it was like 11 or something this is just reshaping the world in our view so this is
our Bay Area Office where we have our AI team main office is over in Norway uh where we do all of our manufacturing development and production Hardware engineering so here at the manufacturing operations and engineering department at 1X uh we have a very diverse team there are people who are responsible for Designing the processes we have a very strong quality team as well that support that system and we create robots that are safe capable and affordable so we can create robots that can live and learn among us a lot of like early inspiration from Ponda
Asimo I think it was like so ahead of his time and kind of still almost this right and followed the field ever since I've always been interested in robotics and AI so as a small kid my favorite movie was The Iron Giant that was not a compliant robot by the way a compliant robot is a safe and soft robot The Iron Giant was very large very stiff very dangerous so I had the fortune of being able to the Google robotics lab really early in 2016 where I spent 6 years at Google brain studying how to
make general purpose robots learn tasks I came to the realization that the algorithms really would scale the Deep learning really was the right way to solve robotics um however the limitation was the hardware in order to really get to the millions of tasks you needed a hardware form factor that was safe enough for the home and also could reach everywhere a human body could reach and you know the most obvious form factor for that is is a human body around 2022 I met met B and we uh realized that if we could merge the compliant
humanoid form factor with a very general purpose AI strategy that could solve lots of tasks that allow us to push into general purpose unstructured environments like the hope so in 2015 we focused on Foundation Technology Building the motors then 2016 we had the first full like two armed robot that could do some stuff allowing us to raise some external capital and start scaling 2018 First full Eve 2020 was All About commercializing Eve starting to run Pilots 2022 we launched Eve 2 where we did the first actual commercial rollouts with real customers and now for 2024
we're transitioning over into Neo which is going to be the first robot where we can fully go into the homes for the last N9 years really been going deep down the rabbit hole of how to design and manufacture every single component that makes us able to create systems that are closer to human so the only way we get truly intelligent robots is to get them into the think about early llms right there was a lot of people thinking like oh we're going to train very narrow wedge approaches where if you train on exactly the data
for the task that you want to apply this to this will create the best model now this clearly proved out to be wrong so the only way you get to truly intelligent systems is in the home where you have this enormous amount of diversity I think that's that's how we learn right we learn by experimentation and observation not just through observation we collect a large diverse set of robot data with our humanoids both Eve and Neo we trade a robotic Foundation model that captures all kinds of knowledge about the world and then we turned that
into a helpful assistant using the same techniques I've been shown be very helpful in language modeling so this is Eve which is uh our first product Eve is about four years old now and this is the first generation of all of our technology so all of these early innovations that we had to do to be able to create systems that are like compliant back drivable safe strong and able to interact with the world to learn Even in our spaces really started here this is what is getting matured into the second generation of the hardware which
is in Neo and where we really feel we can take the final step to be in all of these diverse environments that we have at home to be able to truly learn among people it's all about making sure that there's as little energy in the system as possible so if you think about these classical robotic systems they're often driven by very heavy very high gear ratios inside there's nothing really that's moving any faster than what you can see like the body itself move but in a normal classical industrial robotic system you would have like a
motor and a gear inside here typically spinning 100 times faster than the arm is moving now you have to take that 100 times and you have to square it so now you have a multiplier of 10,000 on the energy of that system and all of this energy has to go somewhere when you instantly have to stop it's all about building systems that are passively safe and like we humans we don't think about this right but if we run around a corner we bump into each other it's not really dangerous for most robots it's roughly similar
to holding a 15 kg kettle bell in your hand while moving so now you can think about like okay if I have a 15 kg kettle bell strapped to each of my wrists and for the legs it's usually like 25 kg and you see this in how robots typically move like when you see a robot walk and it's walking like this actually you would walk the same if you had a 25 kg Cal strapped to your foot because now you need to be very careful right very soft this is also inherently just extremely unsafe this
worked extremely well in robotics for factories when you're in a factory everything is calibr so you can move very fast very precise and whenever you're going to touch something you can slow down right and this is because there's so much energy in the system that if you Collide it that would go really bad and generally in robotics right you call any kind of any kind of contact with the world we call it a collision which is starting with the wrong approach to the problem everything we do as humans in everyday life is colliding like whenever
we're taking a step for colliding or picking something up for colliding it's very hard to know if like am I supposed to touch this or was it an accident so you really need to build into your system how to minimize this energy so that you're just inherently safe you can't trust your sensors on this and that just requires a completely different approach to robotics so what's pretty unique about this is kind of like just like for us humans we have the motors up here with the tendons that are pulling very similar to muscle and this
actually allows you to for example have two Motors here together Drive the two degrees of freedom here so just like for me when I'm doing this and I'm doing this it's the same muscles so you don't have to carry as much muscle to have all this flexibility and that's the same here so you can see like this it's the same due to having this bi inspired approach to the joints where your very low ratio Mo strong Motors are pulling the tendons you get this very compliant safe nature from from a software perspective manipulating bits is
very simple right if you say copy this file from A to B it'll very likely show up in B when you do the copy operation because computers are made very reliable if you contrast this with robotics if you ask a robot to flip on a light switch you don't know if it even touched the switch you don't know if the switch turned on you don't know if it fell over in the process of trying to Turnal light switch if uh just basic sort of operations that robots are asked to do are hard to measure because
you don't really know the state of the world outside of the robot so I would say that the main difference between manipulating the physical world and manipulating the digital world is that you only have control over really what's inside of your robot once your robots are much more intelligent you can start to imagine that they become much more reliable at manipulating simple things in the world like maybe in uh in a few years we can take for granted that robots can pick up objects and deposit them in some other location with extremely high reliability and
that's kind of like being able to copy a file from one location to another in your computer but again like to solve general purpose labor you need the operations that robots perform to be much more reliable than that and we hope that in the limit when robots are as intelligent as humans you can think about the entire physical world as a sort of computer where you can move things around just as reliably as you might move something over the internet to solve the large diversity of tasks that you encounter in the home you need a
lot of data that captures all this diversity this is the same approach that has underpinned technology like chat GPT where you need data across many different types of tasks um not only do you need those tasks you actually need even even broader set of things like you know things from the internet and so forth right so so the analogy in the home might be you want robots interacting with objects in all kinds of way both useful and unuseful once you have all that data you can try to compress that into a very powerful model that
generally understands the structure of that data in language modeling this is achieved by basically predicting the next word and you can think analogously if you're able to transform robot data into a similar format you can also predict what comes next as a way to distill the intelligence about the data in the world into a single model you basically take data around is this good or is this bad and you um align the model to be generally generating good behaviors rather than arbitrary behaviors and so that's the basic approach we're taking at the 1X AI team
our approach to data at 1X is fairly straightforward we collect data from the environments that we want to deploy the robots in I do think that there's many approaches to building a general purpose robot and finding sources of large amounts of data this can come from the internet this can come from human videos this can come from simulation but at the end of the day you're going to deploy this in a home and you need some amount of data in the home so you can't there's no way getting around that we're just going straight to
that approach and we're trying to not be too clever with the data collection and we're just scaling up the the thing that we know is in the test set um and once we have the data that's very closely aligned with the deployment setting we're pretty sure that that data is high quality enough that we can trade things to work in those environments I think the data we've collected on Eve captures a very large distribution of scenarios you might encounter the real world but we get this question a lot actually about like whether we're going to
use Eve data for Neo I do think that in practice um over the next 12 months that data will become increasingly less and less important again it's one of the those um a lot of people try to overthink the data collection problem we think that we built enough experience operationalizing how we collect data that we can pretty quickly catch up the diversity and scale of the data we've collected on Eve pretty quickly on Neo within the matter of weeks so from an engineering perspective it might not even be worth the time to Port the eve
data to the Neo data go to rest so what we have here is a policy that our Android Studio team member Brian Train by himself Open Door go to toilet this is not produced by the AI team but rather a capability that was learned using only data collection and a no code interface that we built here lower suit there you have it that's uh some of the autonomous capabilities we're training on Eve and we'll be bringing those to Neo very shortly um you can see there when it was leaving the door it bumped into the
door but that's totally fine because the compliance of the robot really is making sure that we're designing these robots to come into contact with the world U you know even accidentally sometimes which it will happen building human or robots is a really interesting problem because none of the components that you really need to build this exists so when we say that we manufacture this internally and we have a very deep vertical integration we mean all the way down to like we coil our own coils for our Motors and we literally get copper and aluminum in
and we get Dr it's coming out of the factory and controlling your own supply chain has been incredibly important to be able to kind of control your own destiny as you scale this we start first by ensuring that we have a robust process for the Assembly of the core components so these would be for example our actuators um the motors and the different uh components of the drive units once we finish these we would move to the next stage where we build up the sub assemblies that embody the different parts of the robot uh and
when all these subassemblies are finished and tested then we put the final robot together and then we do a final verification test that it works according to it specifications for Eve we last year peaked out about between 10 and 20 units per month manufacturing wise for Neo we're going to do like 5 to 10x of this in the factory we're building right now and luckily enough we have all of the experience from the hard work of manufacturing even getting that into the market even though the system on the surface might look more complicated since it's
a second generation of our technology manufacturing wise is actually quite a bit more efficient to assemble and manufacture I think working at the intersection of AI and Robotics is really cool because it is the mix of a job that is both challenging intellectually and uh fun like just like when you see the robot actually start to walk everyone claps and cheers right it's it's really exciting to to to see it kind of go get to a working State this is like a perfect career I think for people who are interested in having a big impact
on the world but also working on challenging problems and it's really hard to find actually jobs in the that have all three of these things it's all the cliches right like founding a company and like building it you realize all the cliches are true it's about grit everyone will tell you you're wrong everyone will laugh at you I think the biggest thing you can do as a Founder is to have conviction in your own idea if I look at like companies dying also in robotics it's often because they had this like great intuition about like
this is the thing we should do and then it gets watered out because people just keep beating on them and you're like like for us right oh you can't use tons to drive a robot they don't have reliability durability all of these things that's been tried many times before and then try again and you try again for years and you finally make it work and I think that that's that's the true Innovation right it's a cliche because it's true just like don't give up the next major milestone for 1X is to take all the learnings
we've had on AI and Foundation models and Robotics from our Fleet of eaves and transfer it to Neo and deploy this in the home so this is another level of difficulty from what we've experienced so far because when you're deploying robots in the home you don't have the ability to test new ideas uh at customer sites they really just have to work uh on day one so as a technologist I'm super excited about the possibility that the physical world becomes a computer this would basically be a huge economic unlock in the amount of productivity we
can do of course there's also the benefits to the Aging demographics and I think this kind of technology is really essential for the survival of the species if we're going to keep growing the size of the population and keep a high standard of living for everyone in the planet and then also it's just extremely amusing to see robots struggling and also succeeding I think that really inspires me when the robots actually can do something smart so 2025 will be the year of scaling this 2026 is kind of like still early adopter scaling manufacturing Gathering all
the data to make the systems intelligent enough to be really really genuinely useful 2027 is probably where we start to see that the systems are intelligent enough that we can apply them to other domains so going into everything from like manufacturing to service industry warehouses I think at that point you've hit kind of the sweet spot where these affordable systems are useful enough that the unit economics are just crazy and you're going to be severely Bal knacked by how quickly can you manufacture this 2027 2028 2029 2030 it's going to be all about manufacturing at
scale and how do we get to billions of droids we're figuring out how to process right now but what I can say is that we can manufacture this at a cost of a relatively affordable car and then I'm not talking about a car that someone in Silicon Valley would [Music] buy on next week's episode of S3 we're filming a world first something that sci-fi has been predicting would exist for decades and as far as we know this is the first time it's ever happened in real life we're hoping to film and document many more moments
like this on S3 since the last episode of year 1 we've revamped the show and brought on a small team to produce it at a higher quality level every week over the next 10 episodes you'll notice our continuous improvements in narrative visual quality and a reimagined sit-down interview diving deeper into the technical aspects and strategies Behind These Innovative teams 1X hired us to help create their announcement video but to be clear they didn't pay nor can you pay to be on S3 or impact the way we tell a story about a given company so how
are we making money in the future we're planning to carefully accept sponsorships for the show but what I'm most excited about is our weekly merchandising this is something brand new for year 2 of S3 every week until the end of 2024 we'll be creating custom merchandising for each company we feature you can think of Timeless vintage designs paired with the ethos of building the Future these items will only be available for purchase for 7 days after the related episode goes live then it'll never be sold again our mission and hope is to show that with
the right mindset and determination anyone can work on and contribute to building the future thank you for watching and supporting our work we'll see you next week and until then keep on building the future
