Sunday Robotics debuts Memo, a wheeled home robot trained via data-capture glove from 500+ homes

and measurable. If you're launching a new company, growing, get on adqu.com, get some billboards. Uh our next guest is in the reream waiting room. Let's bring them into the TVPL. We have Nikita from Flexion is also a day job. Is this a day?

No, no, no. We got This is Tony.

Oh, Tony. Hey, sorry.

We got We got mixed around. Uh Tony, so great to have you on the show. Uh I'm sure your 24 hours, last 24 hours have been absolutely crazy. We uh played your demo on the show earlier today and uh we're absolutely blown away. It's it's really uh tremendous progress and we're we're excited to meet you. So before we talk Sunday would love a intro on yourself, background and all that good stuff.

Yeah. Yeah. Absolutely. So excited to be here. So before that I was actually a PhD student at Stanford working on robotics.

So some of the works are like Aloha. You saw like the two robot arms clamped to a table. And it's not just about the hardware but how it learns.

Uh how can we learn from human demonstrations? How can we learn through reinforcement learning and all these things?

And I think last early 2024 is when I have the realization that like you know pumping out more papers and doing more research may not be the most direct way to push robotics forward but starting a company and working on real product is. So this is why I co-founded this company Sunday with Chung uh who is also a PhD student at Stanford and uh which leads to memo uh X1 and all these like uh new advances.

Uh incredible. What uh what has it taken to get to this demo that you released yesterday? Because our our uh I have no idea how much money you've raised uh up until this point, but it feels like you guys have accomplished a ton uh in a pretty uh resource constrained way, at least compared to companies that you're competing with in the sort of like helpful humanoid in the home category.

Yeah, absolutely. So we're we're functioning in a very efficient way and I think as an early stage company we think about as a blessing that forces us to innovate and finding out like these solutions that are orders of magnitude more efficient than like 20% efficient and 30% efficient. Um, and I think a big part of it is also about like the culture and the team and all the people we have that are like really experts and really believes in what we're doing. Um, yeah. Uh, John,

yeah, I I'd love to know some of the

also the chat is mentioning you you forgot to mention you worked at Deep Mind Tesla and and Google. So sort of a non-traditional background into robotics places. Yeah. What what are the key trade-offs? I mean there's a lot of focus right now on teleoperation. Is it is it something just a step in the path towards full autonomy? There are obviously some folks that are jumping straight to uh straight to full autonomy and they say oh we never use uh teleoperation at all. other folks who say uh teleoperation is a really useful tool to pull forward some of the capability. Where do you stand on the issue?

Yeah. So I think teley operation is a really powerful research tool

but it's not necessarily the best tool to get to a product

because if you think about robotics and you know put that right next to autonomous driving right Tesla has like millions of cars collecting data for them every single day and it still took almost a decade to kind of see a light at the end of the tunnel that things are starting to work very well

in robotics if the only thing we can rely on is teleoperation to gather the amount training data it will take like decades for sure because robotics is a harder problem than self-driving.

Yeah.

So the way we think about it is that how can we use human data to train the model? We have like 8 billion humans in the world. Like if you use like 1% of that that's already huge.

Yeah.

So what we designed instead is um I actually have it here is called skill transfer glove.

Uh skill capture glove. Yeah.

That is one to one to man's hat.

Oh interesting. And yes, the idea here is that if you can wear the glove and do a task, memo can also do it.

Okay.

And that essentially decouples this whole like you need a robot to be deployed in the wild before you can gather the data to train the AI. We can train AI just by having people wear a glove and collect data.

Yes. But uh I mean just to go back to the question of like capital intensivity, 1% of 8 billion people, that's 8 80 million gloves. if the glove costs even 10 bucks were back in, you know, you need a billion dollars to get your data set or something like that. Uh,

you don't you don't think Tony can I'm not I'm not saying you can't do it. I'm just saying like like I is there a smoother path here? How many gloves have you shipped? Is is is there a scale thing? And then also I'd be interested to know about like transfer learning. Are you having luck with simulation? Are you having luck with uh there's a lot of video uh just content out there of people doing tasks. Is there any signal that you can pull from just a YouTube video of someone doing the dishes or do you need to simulate something in uh Unreal Engine or use a world model like what are the other tools in the tool chest?

Yeah, I think robotics is at a point that there are so many of these ideas that we haven't converged to this like one single thing which is like pre-training and post- training for LMS.

And the way we think about it is that out of all these methods some will be better than others. Mhm.

And as a startup, we should focus on that one thing that we believe in and build the best system and stack around it. And what we chose was using human data like using gloves to gather data.

Yeah.

And uh actually for all the models that we saw, we of course pre-train uh on like internet scale data, but all the specific behaviors are learned only from the gloves that we make. We don't do tally operation. We don't do simulation and we don't have role models.

Whoa. Okay. Uh then then how do you see the the uh data capture from the glove scaling? Like do you think that there will be 80 million people in five years using this to create more training data or do you think it's a little bit more tractable of a problem where uh at a certain point? Okay. Yeah. It's been a big operation but it's more like 10,000 people that you're employing or something like that.

Yes. So I think this question is more about like for us how can these data generate value

so that we can keep this loop going now

right

and it's kind of similar to the whole large language model space that we need to spend a lot of money into compute but the model itself is generating like tremendous amount of value

and for us we don't need to solve robotics to ship a product that's a lucky part.

Sure.

And in the homes there are lots of like it's one of the few places you can do relatively simple tasks. Yeah.

But give people a huge amount of value both emotionally and functionally.

Yeah.

And and it's much more low stakes tasks than self-driving. Right. So self-driving you said it's a harder you said that uh home robotics is a harder problem earlier if I heard that right. But at least Yeah. But but but at least it's lower stakes and that if you have an air if you drop a dish it it's annoying and you want to avoid that. But there's not like nobody's going to like die.

Yeah.

Yes. It's like the newer start of the pro like of the company is to solve robotics. Yeah.

But we don't need to solve robotics before we ship a product.

Yeah.

So yeah, talk talk about uh timelines.

Yeah. So we've been around for a year and a half and our next milestone is the beta program that will run late 2026. That is when we'll put memo and like tons of them into people's home and actually see how people interact with the robot and what do people want from the robot.

Um and the general availability of memo will be either 2027 or 2028 depending on the progress we made uh through the whole beta program.

Uh talk about form factor. Why not uh why not give it legs? I'm sure you have a a a reason for that and I and I'm curious uh because I think I think people's immediate question is okay I can see how a wheeled system it makes a lot more sense in a lot of ways but what happens if I have stairs? Yeah, absolutely. So the way uh we designed this robot is super safety at a really high priority and the way we define safety is we call it passively safe that if the robot arm and torso is fully stretched out and at that point you cut the power of the robot. Can it stay stable or not?

Interesting.

And a wheel robot is actually like one of the few ways

fall over and crush your dog or or even your foot basically. Yeah. or wreck just like smash the floor, all sorts of stuff. That makes a ton of sense.

And then also, I imagine that there's you can just have more battery power, maybe dock easily, and there just aren't that many tasks that require it. I feel like uh every demo is the I mean, the wine glass demo is remarkable. Um holding two wine glasses is hard as a human, let alone as a robot with kind of odd fingers. But uh just the tidying up use case is potentially underrated because that feels like that feels right around the corner. Even if the like dealing with all the racks and and different spoons and knives and wine glasses, doing the full dishwasher feels a little bit harder. But there's a willingness to pay, at least for me, just to go around the house and pick up the ball that's needs to be in the toy basket and pick up the shirt that's on the floor. Like that's that's valuable. That is actually value. uh if you can get the price right.

What was the uh what was the like key design inspirations? What matters to you with design? Somebody in the chat was asking if you were influenced by Homestar Runner or what?

Oh, yeah. It is Homestar Runner. That's hilarious.

Yeah. So, the way we think about design is we kind of think backwards of what do we want the world to be like? If the robots are ubiquitous, if you need to see it like every single day, what should it look like? and we lean quite heavily towards building a robot that is friendly but also functional.

And these two things there's actually a small overlap in between them. Um so when we designed the robots uh one I think detail that we um decide to do is we do not put camera into the robot's eyes. The camera is actually right underneath its head.

Yeah, I saw that.

Yes. So the reason is that like you're going to make eye contact with the robot. you're going to like look at his face, but if you look at someone's face and his eyes, you see like a camera watching you. It's a little bit creepy.

Oh, interesting.

So, we kind of intentionally avoided that. Um, and yeah,

that's very interesting. Yeah, the Yeah, the design I we were talking about earlier. It feels like it really uh just it avoided like the uncanny valley, the creepiness. Like there's a lot of risk factors when you're designing humanoid robotics right now. We've seen all sorts of them. uh or they can look cool sci-fi but maybe weirder in certain context. I think this one came across very well.

Well, super super excited for you. Thanks for coming on and breaking it down. This is really fun. Thank you so much. If we'd love to be in the in the demo program,

we got flat floors here.

We got flat floors. We have uh huge messes

and we have a team of people that we will make wear these gloves all day long.

And we will take care of we will take care of Memo.

We will

because he's because he's cute.

Yes. We love

and we want to see him win. So, thank you so much. Congrats on all the progress.

Congratulations.

Thank you, guys.

We'll talk to you soon. Goodbye.

Uh, you know what we got to do? We got to get Memo a watch on get bezel.com.

Memo 6,000. Ice it out.

Luxury watched out.

Fully authenticated by Bezel's team of