Building Robots That Can Cook

Rajat Bhageria is the founder and CEO of Chef Robotics, where they are building robots that can do the work of assembly line cooks in commercial kitchens. He's also the cofounder and managing partner of Prototype Capital, a pre-seed VC fund. He was previously the cofounder and CEO of ThirdEye, which got acquired in 2017. He has been a writer for Forbes, TechCrunch, and Huffington Post.

Rajat's favorite book: Steve Jobs (Author: Walter Isaacson)

(00:00) Fundamentals of Robotics and AI
(01:44) Sensors in the Kitchen
(03:18) Motion Control in the Kitchen
(07:15) Introduction to Chef Robotics
(10:15) Designing Robotic Grippers for the Kitchen
(14:01) Computer Vision in the Kitchen
(24:31) The Power of Marrying Software and Hardware
(28:00) Energy Consumption in Robotics
(32:18) User Experience in Robotics
(36:32) Technological Breakthroughs and Future of Robotics
(41:28) Rapid Fire Round

--------
Where to find Prateek Joshi:

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
Twitter: https://twitter.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.498)
Rajat, thank you so much for joining me today.

Rajat Bhageria (00:04.822)
Thanks for having me, Prateek.

Prateek Joshi (00:07.678)
Let's start with the fundamentals. Can you break down the core technologies that enable robots to replicate the precision and speed required in a kitchen? Like for example, assembly line cooks, they do a bunch of tasks with ease. So how does a robot do it?

Rajat Bhageria (00:32.266)
Yeah, it's a very good question. I like how you said it, like the fundamentals. So I think actually the fundamentals are actually basically the same for all robotics and AI. I would say there's like four ideas here. So one is sensing. So part one is you take data from the external world and you make sense, you get it into a computer. So this could be using sensors like cameras, RGB depth cameras, basically, lighters, radars, et cetera. The second part is kind of like the

the intelligence part. So you take that data and you do something with it or you think about it, right? You come up with conclusions. Then there's actuation. Once you have that kind of synthesis of the data, then you do something physically, you actuate the world. And then maybe number four is kind of, you can communicate, right? So you can tell different users about what happened. So broad strokes, I think that's kind of how we think about robotics. And then...

Within a kitchen, I think it's actually quite similar, right? So you have RGBD cameras to make sense of the world, use a lot of machine learning to do the intelligence part, and then different utensils and different actuators and effectors, if you will, to do something. I mean, broad strokes, I think the analogy applies.

Prateek Joshi (01:44.43)
Right. Now, you talked about sensors and if you look at a typical commercial kitchen, there's so many things you have to accurately detect and manipulate, like there's different textures and shapes and sizes. So, is there a special requirement in terms of the sensors you would need to differentiate between raw and cooked states or different things in the kitchen?

Rajat Bhageria (02:13.598)
Yeah, it's a great question. And I think like, perhaps the most important thing is to really, you know, I think like

It's really important to actually pick a segment within any particular industry to really focus on. So for example, we build AI and robots for commercial kitchens, but within that we really focus on what's called assembly and plating. And the reason we focus on that is because it's the most labor intensive and actually expensive part of the process, actually more expensive than prepping and cooking, which is a little bit counterintuitive, but actually true.

So I think broad strokes, like it's important to pick one of those sectors or segments and the sensors you need will actually differ. But within assembly, you know, we use a lot of depth cameras, RGBD cameras. That's, that's really important for us. We also use other sensors like force torque sensors. We use weight scales, but it kind of differs by the application, I would say.

Prateek Joshi (03:18.618)
Right. And another critical component in robotics is the motion control. And obviously in different settings, you need to ensure different types of things that you need to ensure a smooth motion. So in the case of a kitchen, like what are the things that you need to keep in mind when you're designing a system that can do smooth robotic motion?

And also you need precise movements because they're pretty dense. Many people moving around, it cannot just like abjectly do something and physically hurt the next person. So how do you do motion control here?

Rajat Bhageria (03:57.879)
Yeah.

Rajat Bhageria (04:02.75)
Yeah, yeah, it's a good question. And I think there's a lot of different aspects to it, as you said, right? Safety is actually a really important aspect to it. And the reason safety is actually important is because I think within manipulation, there's like kind of two ideas, like on how to actually deploy robots. One is kind of this idea of you fully automate the process. And if you fully automate a process, you can basically use something like an industrial robot. You can put hard guarding around everything and...

You never expect a person to get into that workspace. But the caveat is that if you want to do this full automation of the entire process, you have to have a pretty high level of autonomy. You have to be able to handle hundreds, thousands of different ingredients. You have to be able to handle.

different ways of cutting them and cooking them and prepping them. So it's actually pretty similar to the autonomous vehicle problem, which is like, if you want to deploy autonomous vehicles, you need to have like a really high level of autonomy before you even allowed on the street, on the roads, right?

So that's kind of one idea. The other idea is kind of like more like partial line automation. Let's use robots and humans together. And with this, of course, the benefit is that robots can do the majority of the job, but maybe humans can take care of some of the edge cases, some of the really hard ingredients. And of course, in the early days, as you're building out your training data set, it's really helpful to start off with that second idea. Let's start with partial line automation and let's deploy robots so we get more training data, which then allows us to do the former.

the higher level of autonomy. But what does that entail to your question? What that entails is if you're doing partial line automation, you're working very closely alongside people. So you can't actually use an industrial robot as easily because you might hurt somebody. So we use a lot of collaborative robots, right? And these collaborative robots actually have kind of sensors in each of the joints, which basically will safety stop or protective stop.

Rajat Bhageria (05:55.846)
any time they are above a certain force limit, like 150 Newtons, which is like 15 kilograms or so. So what that means is that it's actually safe to use alongside people, which is really nice. And then more generally, I think, the majority of our company is actually robotic software engineers. So there's a lot of work we do when it comes to the motion planning. And a lot of the work when it comes to food is actually less

Um, let's do.

It's actually more like what we consider like grasping or manipulation, if you will, more than like the motion per se. Like the motion of the arm is actually not as important as like how the end effect or the utensil interacts with the food, if you will. Right. In other words, you don't want to damage the ingredient. You don't want to you want to be very consistent. You don't want to have a bunch of food left over in a pan. You want to place it very precisely. You don't want to spill it in transit. And we think of this more like manipulation, if you will.

is really the hard part, I would say. More than like, you know, taking a robot from here to here. That's actually like, you can use a bunch of off the shelf hardware, off the shelf software to just go from this pose to this pose. It's more like, how do you actually do manipulation dexterously, that's the hard part, I would say.

Prateek Joshi (07:15.49)
That's amazing. And maybe it's a good stopping point to quickly talk about Chef Robotics. You're the co-founder CEO of the company. For people who don't know, can you quickly explain what the company does?

Rajat Bhageria (07:30.71)
Yeah, so I think what Chef does is we build AI-enabled robots to help the food industry kind of overcome their labor shortage and really help them increase production volume and yield. And there is, of course, some context here, which is the food industry is reeling with this really, really tough problem, which is that people don't want to work in the food industry anymore, very, very commonly. Because it's just not.

It's a hard job, right? And so the result of this is that a lot of food companies are running way under capacity. In other words, they could be producing X units, but because they don't have enough people, they're producing 0.7 X units, which of course they're just leaving revenue on the table. Of course that also leads to unsatisfied customers or long wait times, things like this. So...

What we basically do is say, okay, look, well, the crux of that is that people don't want to do this job. Well, if we can build an AI enabled robot that is as flexible, as dexterous as a person, then we should be able to provide a one-to-one human equivalent. If we can do that, then instead of producing 0.7x, you can produce x, right? And so that's kind of the kind of hair on fire problem that we're trying to solve, right?

But of course, there's a lot of ancillary benefits of this. Over time, we can help our customers reduce food waste because robots are very precise. We can help train those workers and help them become robot operators and up-level their careers. And then more on a macro level, this is far out, of course, but on more of a macro level, you can imagine that if food companies keep on reeling with...

lack of labor availability, kind of lack of labor, right? They're more likely to think about an offshore parts of their production, at least parts of their supply chain, which isn't great for the United States. It's really not great for any particular nation. You'd rather have all of your food production onshore, right?

Rajat Bhageria (09:30.814)
So I think AI plus robots can really help with that as well, keeping American food production, manufacturing, service, and supply chain as much as possible onshore. So I think that's kind of how we think about chef and what we do.

Prateek Joshi (09:45.198)
Right. And earlier you mentioned the task of manipulation, which is a, it's a hard task. And especially when you look at a kitchen where you need to be, sometimes you need to be delicate and for a specific culinary task, it requires you to do certain things very delicately. So in terms of the design of the robotic gripper design, what, like what should the robot get right?

What are the things as a robot builder, what are the things you need to keep in mind when you are building like a robotic gripper in a kitchen?

Rajat Bhageria (10:23.958)
Yeah, yeah. No, I mean, it's a very good question, and there's a lot of really important things that matter. Let's just start with some of them. So, perhaps probably the most important is the breadth of ingredients you can do. And this is the crux of everything, because if you're trying to build a machine that just does one ingredient, well, that's been solved 50 years ago. You can get a dispenser, and the dispenser doesn't really have any sensors. It doesn't really have any intelligence. It's just motors that are kind of doing the same thing over and over again.

The thing that's hard and the reason that humans are still doing this job is because it's, it requires some intelligence. So, you know, for example, you take an onion and you cut it different ways. For example, you julienne it versus you chop it. Well, it's essentially a different material and has different dynamics, different kinematics, different physics. If you cut it into quarter-inch cuts, that's different than half-inch cuts.

For example, if you saute it, that's different than room temp, and room temp is different than broiling it. Well, the onion from South America is different than the onion grown in Africa. You get the idea. It's a very highly dimensional space. Now, if you add a little bit of, if you add one supplementary ingredient like oil, then it's a little bit different than water. You get the point. It's a very highly dimensional space. So probably the most important metric is the breadth of ingredients you can do. And how, and the secondary metric with that is like, how consistently can you manipulate those?

In other words, if you were to pick any particular variant I mentioned, can you get within a specific standard deviation and mean away from the target weight with a high probability? I think breadth of ingredients while maintaining precision and accuracy of that as well is important. Then we think a lot about do we damage ingredients in transit? Do we crush the ingredient? If a consumer were to look at the food, would they feel like it's...

tarnished in any way. So we think a lot about that. We think a lot about yield. Yield is basically like do we waste food somehow in the system? Whether we can't pick the food in the pan, or we spill it in transit, or we pick too much, or we pick too little, which leads to waste. Yield basically. Yield is really important because for most food companies 40 to 50% of their cost of goods sold is food cost. So helping reduce yield is actually really important to the ROI of our product.

Rajat Bhageria (12:43.594)
We also think about obviously like engineering related things like reliability. Reliability is super critical, right? We think a lot about throughput, how fast are you going? Because if you're going slow, it doesn't matter. We think a lot about like how usable is the system, because a lot of our operators are like not English speaking. A lot of our operators aren't technical. A lot of our operators actually...

perhaps haven't been to college, for example. So how do you make a system that's this complex piece of software and hardware as simple to use as a piece of kitchen equipment, like a toaster? And then the final part of that is how easy is it to go and just let them run unsupervised? In other words, I think a lot of robotics companies have this issue where they'll make a robot work, but there's more.

robot engineers than there are people that used to be doing that task, right? That are now doing something else. So you didn't really do anything per se. I mean, so we think a lot about, okay, like how unsupervised is this? And how easy is it to use that like, you know, a chef technician can just walk away and it's okay, the robots will still perform well. There's a bunch of other metrics, but these are some of the most important ones.

Prateek Joshi (14:01.75)
Right, right. That's amazing how many variables you need to think of in designing a system like that. I wanna talk about the robot's ability to see. So in this case, computer vision, right? So obviously computer vision, a lot of it is solved, a good chunk of it is still in active research. So when you think about what you can take from the computer vision...

arsenal or the toolkit, right? So what can you take reliably and put it on a robot so that it works almost always? Like what can it do well when it comes to seeing?

Rajat Bhageria (14:44.842)
Yeah, yeah. So computer vision has made a lot of strides in the last, like, you know, call it like 10, 15 years, especially after AlexNet and DeepMill Networks really started to kind of take off. And of course, more recently with generative AI as well. What's interesting about all those cases is that, or many of those cases at least, is that...

The training data set is oftentimes, not always, but oftentimes available. So let's just talk about something like an LLM. I know that's not computer vision that you're talking about, but it's a useful analogy, I think. The training data set is like the open internet. You can just download the open internet and that's your training data set. Well, in our use case,

Surely you can use some models, right? You can use like open source models, but the model itself is actually not, I would argue.

Prateek Joshi (15:57.89)
Thanks for watching!

Rajat Bhageria (16:19.905)
Hey.

Can you hear me?

Prateek Joshi (16:24.386)
Yeah, it's okay. I'll edit this part out. Oh good. So you're about to say the model and then it cut out.

Rajat Bhageria (16:26.815)
Yeah, sounds good. Sounds good.

Yep. Yeah, so I think we certainly use some custom models, and we certainly use some off-the-shelf models. But I would argue, actually, that's not necessarily the hard part. I would say the hard part is really the training data and the data engineering around the training data. So with food, there's no off-the-shelf training data available. You can't just go to the internet and download food manipulation training data. It just doesn't exist. And

It's actually not trivial to make the training data too, because let's just say I want to build training data for like a particular kind of stew. I just, whatever, I just made that up. Well, I can make some stew at our office, but that's gonna be very different than customer one, and it's gonna be very different than customer two, and every customer is gonna have a different stew. And by the way, that stew is gonna change day by day, hour by hour, person by person, you get the problem. So it's actually not trivial to generate training data.

either, which is actually why AI plus robots is hard. There's this like cold start problem. How do you get training data to do manipulation? You have to generate it. And arguably, I think truthfully, the only way to generate it is to truly deploy robots in production. And the more robots you deploy into a production use case, the more breadth and depth of stuff you see. So for us as a food manipulation company, we see more breadth of ingredients. We see things from like, uh, you know,

diced chickens to different cucumbers and different ways of cutting the cucumbers to different sauces and different viscosities of the sauces. Now, how does the sauce change if you add some chicken bits into it? Or how does the sauce change if a little bit more viscous to leafy greens like spinach and kale? So you see more breath. We also see more depth. You see how cheese grits changes day by day. And that's useful, right? Because the more robots you have in the world, the more training data you have. And the more training data you have,

Rajat Bhageria (18:26.431)
the better your autonomy and computer vision gets. And the better your computer vision and autonomy gets, the more useful your product is. And the more useful your product is, the more utilization and use value, ROI you can provide to your customers. The more useful product you can provide to your customers, the more robots they buy, which means even more robots in production, which means more training data. But it also means case studies. If you have happy customers, that means case studies. Those case studies can be useful to get new customers.

which then you have the same cycle. Now you deploy with them and you get even more robots and you get more training data set, which makes your prior customers happy and your new customers happier. And this cycle continues. And I think this data flywheel is arguably the most important part of like, not only the computer vision part, but also the manipulation part. You were talking about the motion planning control. Like, I think they're, they kind of fuse because it's just, it's not computer vision versus the robotics and motion planning. It's kind of like manipulation. And there's so, and hardware is part of this too. Like all three of those need to come together.

to really manipulate something really, really well, really reliably, really breathfully, which without hard coding and doing like a bunch of custom stuff, if you will.

Prateek Joshi (19:33.718)
Yeah, I strongly agree with that view that if you really want to build a differentiator, the only thing that's truly yours is the data that's generated from the usage of your product. Meaning you deploy your robots and by definition, all the interaction data, meaning however the robot's being used, whatever it's seeing, whatever the customer's doing with it, every second of that belongs to you and nobody else and that can become a very nice

loop that compounds over time. And I would say that's like one of the only true ways of building a defensible business here. That's fantastic. Now I want to talk about a key aspect that comes up in a commercial kitchen. For example, if you have a robot in a warehouse, it's normal. It's room temperature. It's nice. It does its thing. But in a kitchen, the temperatures get

Rajat Bhageria (20:12.991)
Yes, yes, yes.

Prateek Joshi (20:33.074)
So, thermal management is another thing, yet another thing you need to worry about. So, how does that impact the design of your robot? What extra things you have to do here?

Rajat Bhageria (20:45.887)
Yeah. Yeah, no, it's a great and it's very astute question actually. Yeah, it's interesting, right? Because kitchens actually, they can go both directions. They can go very hot as you called out, but they can also go very cold because you have cold rooms actually. So it's both cases.

Prateek Joshi (21:01.226)
Yeah, right, right.

Rajat Bhageria (21:08.759)
it is tougher for sure. And there's a lot of dimensions to this. I mean, I think really the crux of that is like really good hardware engineers, right? Who can really spec out and source really good hardware that meets those requirements. The good news is that I think a lot of people in Silicon Valley, especially a little bit scared of hardware, but.

There's a lot of hardware, the idea of hardware, it's one of the oldest industries on planet Earth. Right, if you think like, we've been building hardware for millennia, right? Like a knife is hardware, for example, right? Like a fire is hardware. Like we're very good at building hardware. We built planes and jets and we're very good at it. So the industrial automation community, which has built the Tesla cars you might have or the Ford cars you might have, or your black and decker drill or your AirPods, like,

That automation community, traditional automation community is actually quite mature. Now they don't know how to use autonomy in AI. So they don't know how to make things flexible, but they can make a machine really good. So what's nice about this industry is you can just go to the right supplier. You can say, hey, look, I have these requirements. It needs to work within zero Celsius and 100 Celsius. It needs to be IP69K, which means you can spray it down with high pressure wash down.

It needs to be this sanitary requirement. And you can source the right parts. And sometimes you can't, and you have to make those parts custom. But actually, the hard part when it comes to the hardware engineering is maybe I'd say twofold. Number one is the systems integration. Although those parts exist, the supply chain part is actually not trivial. Like, how do you find the right parts and put them together in the right way and have them talk to each other and marry?

marry all the parts together, the drivers together, the low level software together, right, the firmware. That's actually not trivial. The systems integration is not trivial. So that's part one. And then there's a lot of like end effector and utensil engineering. That's like the other big bucket of hardware we spend a lot of time on. But that problem is not a trivial one because there's ramifications. So for example, like if you're in a cold room, for example, then...

Rajat Bhageria (23:18.723)
your robot can't go at max throughput. Like, you know, like in a normal room, it could go very fast, but in a cold room, it takes some time to warm up. Okay, well, there's some software ramifications there too. So it's not a trivial problem, but the good news is that we can kind of stand on the shoulder of giants, at least from a lot of the off the shelf hardware. And then we can really put our blood, sweat and tears into the autonomy of computer vision, the machine learning, the manipulation aspects.

Prateek Joshi (23:42.29)
Right. Actually, you make a very good point. And historically, going back thousands of years, humans are fantastic at building physical product, like hardware, like physical infra. We're really, really good at that. And you're right, like even in Silicon Valley, the people are...

They're scared of hardware in the sense that, oh, it takes a long time, it needs a lot of capital. But historically, I think enormous returns have accumulated to people and builders and companies who figured out how to do these things well. And then we look at Nvidia and Tesla, like some of the biggest companies in the history of all humanity, right? They're all hardware companies. So that's really good. And actually, I read recently, like this decade, we are entering the era of...

Rajat Bhageria (24:17.081)
Yes.

Rajat Bhageria (24:24.633)
Yes.

Prateek Joshi (24:31.59)
techno industrialist builders, meaning it's like industrialist builders of the early like 1870s, but with a new like technological infusion of like the mini software and AI tools. So intelligent hardware. So yeah, that's very interesting.

Rajat Bhageria (24:37.591)
Thanks for watching!

Rajat Bhageria (24:41.043)
Yeah.

Rajat Bhageria (24:45.875)
Yeah, and I think it's more than sorry to interrupt like this. It's actually like it's really powerful actually. And here's why it's really powerful. If you just look at like just raw numbers. If you look at global GDP.

The vast majority of global GDP is not software, right? It's not bits, it's atoms, right? It's medicine, it's healthcare, it's transportation, it's construction, it's real estate. It's these physical world things. We live in a physical world. So software is amazing. It has high gross margin and scalable marginal cost of zero. I love software, but the marrying software and hardware.

That's magical because you can take all the magical qualities of software. You can apply it to an industry, by the way, that's literally like 20 X bigger. Right. And another, another way to think about it. Okay. Like the labor industry, what is the biggest industry on planet Earth by far? It's humans. It's a labor industry. So, well, you can't, you know, yes, there's, there's some jobs, accountants, paralegals, bankers, sure. Pure software, pure, like software sitting in some data center in the cloud.

can help with that. But the vast majority of labor is in the physical world. So you do need to marry software and hardware to have that level of impact. And I think it's quite exciting because for the very first time in history, like I called out that Alex Net moment, I think for the very first time in history, you can marry hardware and software because of computer vision. Before, computers had a really tough time making sense of the physical world. Like if I have like, if I want to pick up an object,

how do I get that information into a computer or some kind of like controller to make sense of it? It's really hard to do that. The sensors are very basic, but we have this like really exciting time where you have really cheap sensors. Like there's RGBD cameras that are selling for like dollars now. Like, and you can get like very simple, like really good sensors with high resolution, lots of pixels, very cheaply.

Rajat Bhageria (26:43.879)
So that's really cool. We have really good intelligence. Machine learning is really good. You can get your machine learning models running in the cloud. So that's really awesome, too. You don't have to buy all this infrastructure. We have GPUs, and we have GPUs in the cloud. And then as I called it, industrial automation has done a really good job of actuation as well. The hardware necessary to do basically any task in robotics, I would argue.

is more or less existing. Like the thing that doesn't exist is the autonomy level. So I think a bunch, there's a bunch of forces that are like making this from the, making this possible from the technology side. And then also from the macroeconomics market size. And those things hopefully come together. And yes, like we see this like huge boom in software plus hardware and applying software in the physical world.

Prateek Joshi (27:30.53)
Yeah, yeah. And I, that's actually a great point about marrying software and hardware and building a beautiful product that actually works in the real world. And that's truly magical. And we use that everyday, our phones and laptops, all of the electronic devices that we so love and adore all of that. As once we love our beautiful marriages of good hardware and good software, just working and doing the job. All right. So.

Rajat Bhageria (27:45.697)
Yep.

Prateek Joshi (28:00.562)
When you look at a robotics product, energy consumption is another aspect to think about. So maybe a two part question here. One, how do current robotic systems fare in terms of energy consumption? Is that a thing? Is that even a consideration? And two, what innovations or what's happening in the field right now to make a robot more energy efficient?

Rajat Bhageria (28:29.819)
Yeah, it's a very good question. What I would argue is I would allude back to the point I made earlier, which is hardware engineering. Electricity was invented in the late 19th century. So the problem of how do you make electricity really powerful but also kind of cheap?

It's not just specific robotics, right? Like, you know, every industry cares about this. So I think there's been a lot of innovation independent of robotics. So honestly, like there's, there's these really, these huge companies, ABB, for example, that ABB makes power plants, but ABB also makes robots. So I think, again, we're kind of standing on the shoulders of giants. They're like the, there's a lot of industries who like they're like every penny is super important. And those people have made motor controllers and.

kind of power supplies and all these things really, really efficient. And robotics has kind of learned from that. So I'll just give you an example. Our robot module, it just requires a normal 110 AC outlet and it's extremely low power, like on the order of like, I think it's like.

five bucks a day type of thing. It's like pretty cheap, right? From a purely electricity perspective. And that's not because we did any innovation there, to be clear. That's just innovation that not even the robotics industry did, like other industries did. Like heavy engineering industries did, that robotics then used, and then we used.

Prateek Joshi (30:10.13)
Right. That's really good. And also that it's fairly low cost. Like there's so many other problems you should work on. But given, I mean, that's five bucks a day, it's nothing. So it's not like the juice is not worth the squeeze. Even if you get five down to four and a half, who cares? I mean, it doesn't matter. Yeah.

Rajat Bhageria (30:18.744)
Yeah.

Rajat Bhageria (30:25.895)
It's not worth it. So the ROI from robotics is actually, yes, exactly. And the ROI from robots is actually like, it's usually like extremely good, not just from like the first level, but the second level and third order effects, right? Like first order, you're like, okay, well, now Bob can go do some other tasks. That's like first order. Second order is like, okay, well, if I have a lot of these robots, then like they're always gonna show up on time and like, you know, six days, that's second order. Third order is like, okay, well, now I have more predictability.

it can increase throughput. Fourth order is like, I can probably have better quality, right? I don't have as much waste. And it just keeps on going down. So like from an ROI perspective, it's actually really good because it's a win for customers. It's usually like the, even if there's like that $5 in like energy costs, like the...

the savings in waste from an emissions perspective is actually way more important. So like, for example, the food industry, like the food industry is actually one of the biggest contributors to carbon emissions. But if we can spend a little bit of electricity and instead of that, substantially reduce food waste, it's actually like a no brainer. You'd obviously do that. So that's another interesting kind of tidbit.

Prateek Joshi (31:48.846)
Right, right. That's actually very interesting. And when you think about the phrase user experience, and again, in the context of software, it could mean, oh, the buttons are beautiful. It was easy to log in and oh, Slack was so nice and easy when it came out, that's software. Or maybe iPhone, like, oh, the unboxing, it just works. It has everything, it's integrated. It's super easy to use. So when it comes to robotics, especially B2B, where you...

your customers are people who run these kitchens, you're interacting with chefs and kitchen staff, how do you think about user experience for a product like yours?

Rajat Bhageria (32:28.327)
Yeah, it's critical. It's like perhaps one of the most important things. In my initial paradigm, I said communicate, right? And that's a lot of robotics, and we've actually missed that. I think that's like, stands kind of like shoulder to shoulder with sensing intelligence and actuation. Why is that? I think in robotics, there's a bunch of people who build cool stuff. They build technologies. It's just like in software, there's a bunch of hackers and hackers can build cool tech. Some people build products.

Right? And products say, okay, well, you solve some problem, right? There's very few robotics companies, I would say, and this is true in B2B SaaS as well, but there's not as many companies that provide a true solution, right? And true solution is truly like an equivalent system. So for us, to provide a solution, like things like, the questions you've been asking have been good, like it's like, it's like,

Okay, does it work in a really hot environment, a really cold environment? Is it really easy to clean, for example, in a food environment? Sanitation is really important. I alluded to the idea that many of our users, first of all, users are different than the customers. The customer is this kind of like corporate executive, right, but the user is usually a Spanish speaking person on the line, right? And our users, you know, we have a bunch of like,

super smart computer vision and robotic software engineers who are from Silicon Valley and worked at all these cool companies. But now we need to make a system that like is super complex and has hundreds of thousands of lines of code work as simple as a toaster oven, right? So that's super critical because we can build this awesome product, but if the product isn't used, why did we do it? There's no point. And...

And there's so many different users, right? There's like the line workers, there's a line leads, there's sanitation, there's kind of the people who are refilling, there's all these different like users and they all have different requirements. So I do think that's really important. And honestly, it's not like, the fundamentals are still the same as like user experience for like Slack, but of course you have the hardware component too. So like industrial design is also important.

Rajat Bhageria (34:44.731)
We think a lot about, okay, well, like how easy is it to move the robot around? Like, is it super heavy, for example? That's like, if you're like...

like, yeah, how heavy is the thing? Like, or the HMI, like, you need to be able to interact with the robot, like, how tall do you, like, how tall is the average user for Chef? So that, like, you don't have to look up or look down at the dashboard. Or it's pretty loud in these environments sometimes. So how, like, if something is failing, like, say there's like an issue or the system's out of ingredient, the refill scenario, how does the robot communicate with you that it needs support?

Like these things are actually really important because if you build a product that the users don't trust or they don't like, or they feel is preventing them from going home, because of course they're not being paid a ton, then they're going to push it aside and you're going to be out of there in a week, basically.

Prateek Joshi (35:40.91)
Right, right. I think that's a fantastic point. And you're right, when it comes to hardware, the engineering part of it is so hard, it's hard enough that people focus all the energy to towards that. And then the most important thing for the customer, which is like, how am I experiencing this product? That is, you're right, that's kind of semi neglected. It's almost like an afterthought. It's like, hey, look, we made this hard thing work. What more do you want? But I think it's a big missed opportunity.

Yeah, you're spot on here. All right, one final question before we go to the rapid fire round, and that's around what technological breakthroughs that are happening in robotics right now are you most excited by? And also, what can we expect in robotics in the next 12 months?

Rajat Bhageria (36:32.583)
Yeah, no, I think it's a very good question. There's a lot of innovation happening in like every aspect, which is pretty cool. I think. So.

Rajat Bhageria (36:48.327)
I think for sensors, I think that we talked about the four kind of pillars, right? So the sensors are getting better and better, right? Lighter is getting cheaper and cheaper, which is exciting because it's such good resolution. There's solid state lighter. I think that's really exciting. Cameras are continuing to get better. Death cams are continuing to get better. So I think that's, I think that's just going to continue. And honestly, like I said, like the hard thing is like, like the hard thing is intelligence, but if you have better data to do, do something with.

Then of course, your outputs can, it's like the garbage in garbage out analogy. Like if you don't have good data as an input, you can't do much with it. So I think sensors getting better is pretty cool. And the cost of sensors as well, like lighter exists, but it's just expensive, for example. On the intelligence side, of course, I'm guessing this community knows a lot about that, right? There's a lot of new algorithms. There's a lot of new, I think there's a lot of good MLOps software too, right? Like that's important, right? So I think that's really good. There's a lot of compute improvements. I'm excited about the compute improvements, honestly, because like,

Obviously, we had a step change from going from CPUs to GPUs, but now we have even better hardware, like machine learning, like model, like hardware that's built for specific things, like ASICs and things like that. And so I think that'll be a big step change too. Like how do we have just better compute and better chips to do things with? So I think that's probably the, I mean, that's of course, like, that's where we're seeing the most innovation right now.

I think the actuation side is kind of like a consistent thing because it's the oldest industry, right? Like we call it like industrial automation has been around for a while. Like it's the oldest industry. So it's kind of like consistent. I don't think there's any fundamental, like new breakthroughs. There's a lot of like research happening, of course, into soft robotics and things like that. But I'm probably most excited about the kind of sensing, but of course, mostly the intelligence side, which is the hardware of the intelligence and the software of the intelligence. That's where, of course, there's the most breakthroughs. And I think that's where it's the hard problem.

The reason robots don't exist, I mean, people often ask me, why don't we see robots everywhere? It's intelligence, it's not hardware. People think robotics is a hardware problem. It used to be a hardware problem 50 years ago when, you know, Ford was trying to automate, but now it's purely an intelligence and software problem. So the same things that are gonna help get Dolly working, the same things are gonna help get robots in the world. So I think what I would hope, what I would like to do is,

Rajat Bhageria (39:07.623)
I would like to see a lot more kind of...

Rajat Bhageria (39:15.683)
I think like what's exciting is that Chachi PT and OpenAI has kind of like inspired people to really like it's kind of like a like the next kind of like boom of AI and I think it's really cool, right?

But I am hoping that more and more people realize that AI in the physical world, and in other words, embodied AI, is perhaps even more powerful. That's kind of the singularity, right? But if you think about what people are scared about, but also what people are excited about, it is really AI and robots in the physical world. So what I'm hoping is that there's a few success stories in robotics and AI that inspire people. ChachiPT and OpenAI was the success story in AI recently.

Prateek Joshi (39:41.631)
Thank you.

Rajat Bhageria (39:58.367)
There hasn't really been a big success story in AI plus robots just yet. Locust Robotics is one that's doing fairly well. You could argue Tesla is kind of like a robotics company with the vehicles. Not optimists, but the vehicles are kind of robots. But hopefully, Chef and there's a couple other companies like Chef that succeed massively. And just like we saw a bunch of people come after Uber and Airbnb for the marketplace economy, hopefully, if we can set a

If we can succeed, then we can inspire a lot of other people to succeed. That's actually kind of our mission. Like it's not necessarily just like Chef successful. Our goal is to like accelerate that rent of AI in the physical world, which also of course means it's just not, we can't have that much impact as a company alone, but if we can be successful, then thousands of other.

Prateek Joshi (40:31.527)
Right.

Rajat Bhageria (40:48.911)
Young founders, engineers, operators, investors can also support those next generation of robotics and AI companies. And that's how you kind of really accelerate AI in the physical world and embodied AI.

Prateek Joshi (40:59.002)
Right. That's amazing. And it's so exciting to be living in that moment where that transition is, I'm obviously rooting for it. And I think it's a transition that needs to happen because I think nothing creates value like an actual physical economy and physical product. So, all right. With that, we're at the rapid fire round. I'll ask a series of questions and would love to hear your answers in 15 seconds or less. You ready? All right.

Rajat Bhageria (41:27.521)
Yep.

Prateek Joshi (41:28.722)
Question number one, what's your favorite book?

Rajat Bhageria (41:33.831)
I think probably the most impactful, I think there's like favorite and impactful, like probably the most impactful book has been like Steve Jobs biography, which sounds colloquial and trite probably, but like that's the thing that kind of inspired me to kind of do entrepreneurship and tech probably from a pretty early age.

Prateek Joshi (41:49.718)
What has been an important but overlooked AI trend in the last 12 months?

Rajat Bhageria (41:55.335)
I would say A.I. in the physical world and body A.I.

Prateek Joshi (41:58.066)
Right. What's the one thing about robotics that most people don't get?

Rajat Bhageria (42:03.835)
It's a software problem, not a hardware problem. And the hard thing is software. Like 90% of our team is software, not hardware. Yeah.

Prateek Joshi (42:07.698)
That's amazing.

Prateek Joshi (42:12.254)
Right, right. All right, next question. What separates great AI products from the merely good ones?

Rajat Bhageria (42:21.223)
I think it's just a long tail, right? I mean, and it's an obvious fact, right? But like, or maybe it's an opinion, but it's obvious, right? But like, something that just like, you call it the iPhone, the reason people love the iPhone, it just works. It doesn't work 99% of the time, it works 100% of the time. It just works. That last bit is fricking hard, but that's what separates great from good.

Prateek Joshi (42:42.31)
Right. That's actually, that's a really good perspective. All right, next question. What have you changed your mind on recently?

Rajat Bhageria (42:51.027)
I've changed my mind about a lot. What's something I can call out here?

Rajat Bhageria (43:00.211)
I mean, one thing that just topical, I used to think like from a team building perspective only get really, really senior folks, right? Like very senior, like I want like staff level engineers only and I think that's important because, and the reason for this is like they really set the architecture well and they make sure you're going in the right technical direction.

Right. But I think there's something about just having a balance with also people who are just hungry and smart. And a couple years at a college, a couple years of experience, but just hungry. And that combination is really powerful. I used to just think just the former, but I think now I'm understanding that the latter is also really important. And the marriage of the two from a team building perspective is also useful.

Prateek Joshi (43:39.962)
Right. What's your wildest AI prediction for the next 12 months?

Rajat Bhageria (43:47.952)
Um, well, I'm hoping that there is kind of, yeah, a couple companies and AI plus robots that they're really succeeds. And there's this kind of chat, GPT moment, if you will, for robotics. And we see just the amount of hype for AI plus robots as we do in chat, GPT and LLMs.

Prateek Joshi (44:05.75)
Yeah, right. Final question. What's your number one advice to founders starting out today?

Rajat Bhageria (44:16.639)
It's been said so many times, so it sounds kind of like repetitive almost, but I think the number one predictor of success is just not giving up. I think if you just like don't give up and you're like, if you're willing to just like, I think you have to like admit reality that if your product's not working, your customers don't want it, you should admit it. But then like if your customers want something, like they tell you they want it, then you just keep on going and you just don't die. And if you just don't die long enough, you'll be successful.

It sounds so simple, but just don't die long enough and you'll be successful.

Prateek Joshi (44:49.154)
Right. That's amazing. Rajat, this has been a brilliant discussion. Love the depth of the robotics knowledge. Obviously you've done it and it shows when you talk about the various facets of building a robotics product and also a robotics business, right? So it's a different part of what it takes to succeed here. So thank you so much for coming onto the show and sharing your insights.

Rajat Bhageria (45:13.811)
Yeah, thank you so much, Prateek, for having me.

Prateek Joshi (45:16.222)
All right.