He's Building $600M Robots That Folds your Laundry

imagine this you walk into a laundromat

and a robot has been folding laundry non stop for 24 hours

that's what Jason and his team is working towards

fully autonomous single task commercial grade robots

and they just raised $120 million in order to get there

our goal is to power the future of the physical economy

getting these robots to be very robust and uh

can sustain a long duration of like

actually doing a task is a technical barrier

that hasn't been really solved before our work

for what are the right problems that you should have your robots solve

the bottleneck for useful robots is AI and software

these humanoids right now

they're not actually very useful

the cost and the hardware readiness is a big factor

the best way to succeed is to build research and

product at the same time

quick thing before we get started

we have a lofty goal this year of hitting 1,000 subscribers

in order to help more people build really great companies

so if you enjoy the content

learn something new the best way to support us is by subscribing

okay let's get into it

today's guest is Jason Ma

one of the brightest minds in robotics

he's the lead author of multiple award winning papers

recognized internationally for his research

he had offers to return to Google

Deep Mind and Nvidia Meta

but turned them all down to start Dyna Robotics

and now instead of chasing SCI fi human humanoids

robots that look and move like humans

he's building robots that actually work

folding towels stacking

packing in the real world

so Jason thank you for joining me at Founders in Motion today

uh thank you Thea

thanks for inviting

really happy to be here and nice catching up with you after so long

yeah this is an interesting change of scenery for us

Jason in very plain English

what is Dyna Robotics building

at dyno we're building general purpose robots

that power the future of physical economy

so the way I think about it is that we're developing AI

powered robots that can do any task in any business or home scenario

and to start out we have deployed our robots

uh using AI in many different scenarios such as restaurants

gyms fitness centers

etcetera and our mission is to make this robots and our AI

model as general as possible

so you can basically do any task that you wanted to do

so when it comes to robotics

there are kind of typically two approaches that people usually

you usually go for so the general purpose

everyday kind of companion

supportive robot or the very fancy humanoid um

move and feel like humans

so why did you decide to start with everyday tasks like

folding towels

instead of going after kind of the overarching humanoid dream

yeah so our eventual goal is to develop a robot that can do any task

right so to that end

perhaps at some point we'll venture into humanoid

robots but it is our belief from my past experience and also research

that human

robots as hardware is not currently mature and it's way too expensive

but I think at the present moment

the bottleneck for useful robots is AI and software

so we decided as a company to first focus on off the shelf hardware

so basically these are robots that you can buy off the shelf

couple thousand dollars

and then we develop the AI on top of it so it can be really useful

so uh

you might have seen some of her demos

that robots can already fold napkins

fold cloth uh

do packaging at very very high success rate and robustness

and uh

I think that's a stepping stone towards developing like the more human

like you know

form factor if you will'cause

you know if you look at these humanoids right now

they are not actually very useful

and you could only see them behind a screen

instead of actually seeing the robots in action in front of you

and the cost and the hardware readiness is a big factor

yeah yeah

that's super fair

and so folding laundry seems like a very simple task for humans

um why is it actually so hard for a robot to do first day

I think folding laundry is also uh

fairly challenging for humans

for one it's very mundane and tedious right

like I don't you know

I don't like doing my laundry and

I think for cloth garments

they come in all different shapes types

you know shirts

jeans long sleeves

jackets and what's really hard about robots

why is this hard for robotics is for a couple of reasons

so one uh

just taking a step back right

traditionally you have robots as automation tool

you see robots in factory

and what happens is you pre program robots exact sequence of motions

right so for instance

you want the robots to package into like a box of

you know things you want to deliver

it's a pre program sequence of motion

but that's actually

very difficult to do for something like folding laundry because

you know your clothes is like deformable

right the shapes are always different

they're not in like perfect

you know rigid shapes and states

so it's really hard to preprogram a sequence of motion to fold cloth

for you

and what that means is you really need to develop

you know using generative AI tools

but just like language Model

Chat GPT can really do anything based on your language command

you want to develop AI

models that can control robots to do very fluid and dynamic motions

just like how human arms are very fluid and dynamic

and do many different motions to fold clothes

so that is why this task traditionally has been very difficult for the

uh preceding paradigm of robotics where you pre program the motions

and now it's actually more amenable for the new wave of robotics

which is learning robot actions through like data right

you learn from data yeah

yeah and if we take a moment to dive a little bit deeper into

um into the technology behind all of this

I guess at a high level

how do you think about training these generative models to help

robots recognize different

outcomes in

in the real world or a different

yeah in the real world

yeah so the way we train it is very much like how

you know humans interact with the physical world right

so you know

like for us you know

we have our eyes that perceive the world through vision

through perception and then we use our brain to turn these sensory

you know image inputs to our eyes into like

you know actions that our arms and our legs perform right

so very much you know

for training these robots

it's like

we have cameras on the robots that perceive the physical world

and then we train neural network

um to output actions are like actions to command their robots

you know joints and uh

you know different motors on the robots

so the way to do this is you know

you collect input output pairs of images in and uh

robot joints out right

so you can basically uh

control the robot to fold clothes many

many times to collect data

and then then that data gets fed into a model

which then can control the robot autonomously

so you're essentially kind of building your own general purpose data

models and datasets in order to train these um

folding laundry robots yeah

basically yeah

but you see the paradigm here is fairly general right

yeah

it's not really specific to laundry folding

so if you have data that you collected for

let's say packaging

for cooking a meal for cleaning your bedroom

then the robot using the same algorithm

can train models that can do these tasks

and what's really interesting is that you can actually just combine

all the different data sets for different tasks into one model

and that model becomes very powerful

right just much like language models today

it's not just one model that does one particular language task

but one model that can chat with you

write code for you and do many other things

so that's like the vision that we're also trying to do for robotics

but there's of course

a lot of nuance you know

which we can get into later as well

yeah yeah

super cool

I think I read on your research somewhere that there's a high adapt

adaptability from use cases from one to another

so the next use case should be trained at

at a lighter weight and then continuously forward

so you ran this super cool 24 hour napkin test

nearly 800 folded with 99% success rate

for someone who's not technical

why was that such a breakthrough

the reason why that result was very difficult and very impressive

in the uh

research community robotics community

is that you know

we have seen you probably have seen a lot of demos recently

of robots doing cool stuff

but typically what happens is that those demos are very brittle

it took many many shots to get even one video that works very well

so basically robots are now at a place you can shoot fancy videos

but actually getting these robots to be very robust and uh

can sustain a long duration of like

actually doing a task is a technical barrier

that hasn't been really solved

before our work

especially for these highly dextrous

you know complex manipulation tasks

such as folding laundry folding napkins

right

and so that is why this result was very different in that it's OK

it's a result where we across a span of a week

we shot many

24 hour video of the robot just continuously folding napkins

non stop for 24 hours

hundreds of napkins folded without much failure at all

and that was very different than prior works where yes

you can get a robot to do a laundry folding demo

but the success rate is often like about 70% or 80% success rate

so that means if you try to fold 10 t shirts

you might only succeed eight times

so that's good enough for a demo

but it's not good enough for actual real world deployment

right cause

you know if you imagine your robot only succeeds 8 out of 10 times

that would be a very frustratingly bad robot

right so I die

now we are very much focused on not just putting on fancy demos

but actually developing the AI technology to power

like what actually can work in the physical world

yeah yeah

and for anyone that is a little bit confused

it's okay

cause I printed all these terms in case Jason ever mentioned them

but dexterous manipulation is basically

just teaching robots to use their hands

as flexibility as we do

so folding gripping

managing weight twisting without breaking things

mm hmm and okay

while you were training

the robot to get to this very high level of success rate

what's like

one funny or surprising failure that the robot had

before it finally nailed the test

for the napkin example in particular

what was very very difficult is like actually

so what happens at a real restaurant is like

they'll ship you a stack of

they'll give you a stack of napkins

for which you have to fold the napkins one by one

so the robot initially would make the mistake of like pulling out many

many napkins from the stack

right and then now you have a whole big mess on the table

where tons of napkins are at the center of the table

and then where you're only supposed to fold one

so that was a very tricky scenario that the robot got into early on

but uh

you know

we were able to basically train the robot to handle those scenarios

so whenever it grabs multiple napkins

then it would put the actual ones back onto the stack

but then that creates a pretty messy stack on the side

so the robot then had to figure out how to deal with that situation

but eventually our robot was able to

get around all these tricky scenarios and become very

very good but just from this example

you probably realize you know

real world physical AI like embodied AI

you know

basically teaching robots to do things is actually very complex

if you handle one scenario well

there might be other scenarios that you didn't expect

that the robot has to handle

that comes up so it's kind of like

you know walk in mall

you know like

just cause you handle one scenario

doesn't really mean other hand scenarios are handled very well

which is why physical AI is often not very

very challenging and as our model got really

really good another failure case is it pulls napkin too fast

so the napkin just slipped off the table

but besides those scenarios

yeah the robot was pretty much like 100% in folding a napkin

that's present in front of it

there's so many different edge cases in the human world to deal with

and yeah embody AI here is just AI inside a physical body

so like a robot arm that can see

decide and act in the world

I wouldn't say I'm technical

but like I studied a little bit of technical stuff

so like yeah

this is really fun for like the nerd in me

um so

I also wanted to talk about that approach a little bit

so Dyna 1 is both the arm

so the the body

the actual physical arm and the AI

the brain of it

so why is it so important to kind of build them in conjunction almost

rather than just like building

like a general purpose AI brain

that can be applied for any different type of physical embodiment

the way we think about this is that physical AI is extremely hard

especially if you're interested in pushing real world performance

right so if you think about like pushing real world performance

then it's a matter of both the software and hardware

like if your hardware like always breaks

your robots are just not good enough

then you can't actually run a model

an AI model on the robot for 24 hours right

so previously when we when I was doing research in the lab

we often ran into the issue that the robots would break

after a couple hours or like you have to maintain it

or the robot would start overheating

hahaha after like five or six hours

so it's just even physically not possible to run for 24 hours

so from that simple thought exercise

it's probably uh

more clear that in order to get to any real world performance

you have to do a software hardware

uh co design

co iteration but I said though

you know in Dyna 1

the arms themselves were something that we bought off the shelf

but we had to do a lot of things on top of it to make it more durable

or

more uh

you know just has higher endurance

so we could even try running a model for 24 hours

and during the early days of the research

it's certainly the case that the model would do something like

more violent like maybe it would hit the table very hard

so that makes it even more important that the hardware is good enough

but what I find interesting is that traditionally

robotics would have a lot of safety features or safety layer

like you program the robot to be safer

but in the AI age what we found is that the robots are

the models are more intelligent

he actually just does more

you know dexterous

smooth behavior so

it's

much less likely to even do the unsafe behavior in the first place

so uh

yeah that made hardware

reliability also just a much simpler problem than before

one point that you mentioned

I thought was super interesting is that like

even though you bought the hardware

um off the shelf

it still went through a lot of iteration to ensure um

to ensure that it could even operate in 21st for 24 hours

which kind of double which kind of double ends the point of like

maybe hardware is the bottleneck for humanoids

and that kind of more advanced application of robotics

why did the team choose kind of an application like folding laundry

folding napkins as the first test case

use case to play with

one is that like if you want to use a more AI

like data driven approach to train robots

then it's very important that you have a lot of data right yeah

and uh

folding clothes

folding napkin is a scenario where you couldn't really like

break the object that you are like teaching a robot with right

like you know

clothes is like napkins are soft

so like you couldn't really like mess it up

but yeah there were other applications we looked into

like for instance

like loading dishes right

but there

the safety risk is a lot higher if the robot messes up a single time

like it drops a dish then the dish break

to advance model capability

we start out with some task that has the feature where like

you can just have the robots practice try many

many times

and like close folding laundry was like the perfect scenario

because once you fold it very nicely

you can just like disturb it

so the cloth gets like crumpled again

so you can have the robot practice again

if you will right

so that was from a technology perspective

very appealing but from the business side

it's also the case that there's a huge demand actually

like we don't even think about folding napkins in restaurants

but if you go to like I don't know

like Cheesecake Factory Applebee's of the world

there are just so many like

napkins that need it to be folded in the back

office all the time so it's almost like a full time employees job

like their whole job is to do that

so we thought there's a huge need to like

do this kind of tasks to get started

yeah yeah

I mean I think folding napkins is a universal experience

across all restaurants yeah

um and also even folding laundry is such a pain too

I am also not a big fan of it

you've moved from control tests to doing some kind of

pilots in the real world

so when you move from like a lab setting to say a local laundromat

what new challenges pop up

what we didn't realize is that in the office

first of all like we have air conditioning

so the room was like kind of cool

but like in a lot of these real world scenarios

you know you do not have control of the temperature

so like overheating becomes more severe

you also don't have good control of the network

laundromats don't necessarily have the best Wi-Fi right

so if you think about like running models over cloud

then Wi-Fi becomes a bottleneck

where if you're doing some data collection on site

then uploading data also becomes a lot slower

and there is the operation challenge

like how can we trust to put a robot in your real customer site

and not worry about something goes catastrophically wrong

like for instance like the robots folding napkins in the back office

but you don't want it to like catch on fire by accident

that's the hard part of robotics

it's not just like getting AI to work

you also have to like make sure the deployment flywheel goes smoothly

for a business owner like how does the math work

how are you kind of thinking of pricing these robots

and when do they actually

pay for themselves

uh a lot of these businesses are quite price sensitive

or like they're operating

you know it's a spectrum

but let's say like you know

restaurants are perhaps low margin businesses

so we realized that in order for the economics to make sense at all

like your robots have to be somewhat cheap right

so this is also why like when you were mentioning humanoids

we don't think it's ready because any humanoids you can buy

or first of all you can't buy that many

but the ones you could buy are also in the orders of like

tens of thousand 20,000 if not more right

but our robots are like couple grand each

we do like a robotics service

business model so we don't like actually sell the hardware

we just rent the hardware out to different customers

several grand a month to rent a robot

so it's actually like on par

if not cheaper than like typical labor cost in the United States

when you think about the next application for dyno um

where what is the team kind of actively testing with right now

our goal is to power the future of the physical economy

so basically any task that we think is extremely mundane dull

dirty dangerous for humans to do

we're looking into it looking into many different markets

so you know of course hospitality right like hotel

restaurants where discovering use cases

then you know the laundromats of the world

the uh logistic warehouses of the world

the most important thing just like develop a general recipe

so we could actually just

deploy our models to any scenario that we want

moving into a bit more about your founder journey

so you could have gone back to Deep Mind and Nvidia or meta

and if you only knew the eye popping number that they gave Jason

but I don't think I can say um

but instead you chose to risk your founder path

what made you take that leap

the best way to make an impact in robotics is at a startup

robotics is an extremely complex problem

it's not just a software problem

so you wanted to have the freedom

the velocity to do all the things required to get robots to work

and I firmly believe that getting robots to work requires

you to actually deploy the robots in realistic

real scenarios right

and I think that's just not possible to do at a big corporate lab

where you know

robotics is like a research problem to them

but not necessarily something they want to solve right away right

like all the companies you mentioned

their core businesses are not robotics

and their core AI is not in robotics at the moment right

you know they're developing big language model

what not to you know

power their platforms right

competing with you know

let's say open AI anthropics of the world

so robotics is more like a research problem for them

but I felt that

it's possible to actually get robots to work in the real world

with the right team the right mindset

and you know

there's enough funding that's willing to fund this kind of startup now

so I thought it would make a lot of sense to start Diana

so you've always been kind of on the more research side

so moving into start up and more of the business realm of things

what's one lesson that you've Learned that really surprised you

for robotics in particular

it's very important to have a good taste

for what are the right problems

that you should have your robots solve

there are so many robotics companies in the past that perhaps picked

a problem that's too hard or like

too costly so it's actually very

very hard to penetrate the market and to succeed as a business

and that's also and the hardware is also just innately expensive

so there's been a history of

you know like

you know like

venture capitals don't really like funding hardware companies

for that reason

in research

the most important thing in my opinion is also research taste

now how do you pick a good research problem to work on

what happened in the past is

a lot of robotics company went really deep in a vertical

so all their solutions

their software hardware stack becomes very specialized in that domain

so you cannot just go from one domain to the other so that's

the kind of thing we don't want to do at Dina

what is like

your two sentence

pitch for why Dina's different from other robotic companies

at dyno we do both the research and also the product

and I think

the best way to succeed is to build research and

product at the same time

right

so you can think of let's say like

you know chat

BT right

it's a product that you can use right

but there's also a lot of research that goes behind it to make it well

and I think the feedback loop from product to research

and from research to product is what makes uh

you know the existing AI products on the market like so good and uh

so sticky so you looking to robotics

there are companies just purely trying to take some

existing technologies and turning into a product

but I think right now existing technology just as it is

is not ready to get to general purpose like manipulation

right like using arms

hands to do things but at the same time

if you're only focused on research

then I think sometimes you are focusing on problems

or like tasks that

may not really be representative of the real world tasks

right

because there's always a gap between research and actual deployment

so as a researcher you know

certainly in my research papers and all that

all the tasks I were solving are like

toy versions of the real world tasks

but at dyno we go straight to like the real version of the task

so we know that if our research is good enough to solve those tasks

we can immediately

we turn them into product

that feedback loop makes our technology and business much more solid

research is always theoretical

like the name of it is about being very theoretical

so yeah um

iterating it within deployment

and within testing is very important to make it actionable

throughout this journey

what has been the most valuable lesson that you've Learned

something that you really wish someone had told you earlier

building a startup is actually quite hard

yeah yeah

really I think

I think it's been a year

I think it's definitely harder than like when I did my uh

you know research for PhD

cause in PhD I was really just minding my own business

my own research so like

there's not that much else have to worry about

besides the existing research

for a company like dyno we were

really trying to develop like Full Stack Robotics

right going from hardware software

AI everything

then there's always many things moving at the same time

not only just excel at what I already do

which is AI research but also coordinate

collaborate with many different teams and to make sure we always

being able to prioritize the most important things in the company

and allocate resources accordingly

and I think that's definitely like a shifting gear

compared to previously when I was just doing my research

shifting gears from like an independent contributor to

like a leader is always yeah

very hard um

okay so Jason

the way that we like to end these things is

we like to play a quick game

so sure since you're on the frontier of robotics

I want your predictions so in 10 years

who will be doing these jobs

robots or humans

flipping fast food burgers

robots delivering packages to your door

robots or humans

I think you get to choose or you should be able to choose

teaching kids in classroom

uh humans walking your dog

uh humans or robots

I have seen videos of robot dogs walking like real dogs

performing surgery in a hospital

humans surgery is where like safety is so important

and I think the precision required

you know there's a lot of research papers on surgical robots

but I feel that this is one area I would be like very careful

so yeah I would still trust the human surgeon than robots or maybe

you know maybe Dina will solve it one day

we don't know we'll see

yeah we don't know

I mean I'd also think it's like depending on the application right

thank you so much for coming on the show

I really enjoyed the conversation

Learned a lot more about robotic

yeah thank you for hosting and having me

that's a wrap if you like this episode

please hit the subscribe button

it helps us bring on more awesome guest

level up production and drop new series you wanna watch

see you in 2 weeks