This is a transcript of Vibe Coding Is Coming for Robotics | 002 with Dan Lambert, Founder of Peppy. Please note that the transcript has been lightly edited for readability and may contain errors.

00:00 Why robots, why now?
01:27 Turning 10 hours of robotics setup into 10 minutes
02:35 Replacing robots like smartphones
04:18 What is holding general-purpose robots back?
06:09 Peppy’s ~$1M pre-seed round
08:27 What can a $5,000 robot actually do?
09:41 The sim-to-real gap and why robots need to fail
11:57 From a plain-English prompt to physical robot movement
15:32 NVIDIA’s role in the robotics revolution
17:59 From experimenting with robots to building Peppy
20:41 “Lovable for Robotics” and vibe coding physical machines
23:06 Peppy’s December 2026 launch and the robot app ecosystem
26:38 Should robots be allowed to use knives?
28:30 Do robots actually need to be humanoid?
31:15 The moment Dan became obsessed with robotics
33:44 What robotics needs next

Jared S. Taylor (00:00)

It's a good time for robots. Why robots? Why now?

Dan Lambert (00:04)

Okay, so a few things have changed in the industry.

I think one big one is that NVIDIA's newest chipset, the Thor chipset, is eight times faster than the previous chip. And so it's the first time that we can see robots responding in real time.

It's the first time we have this brain that's super powerful that can actually interact and handle things.

The second thing is robots have gotten a lot cheaper.

I have a robot here on my desktop. Actually, I'll show it to everybody. I have a few of these guys.

Jared S. Taylor (00:37)

That's cool.

Dan Lambert (00:39)

That is a $5,000-and-under robot, and it's getting cheaper.

So this is the first time that I think it's financially viable that you can have a robot that's doing a lot of tasks at a reasonable price point.

If it breaks down, it's kind of like a laptop. You're getting a new one.

And then I would say number three is that up until now, robotic coding has been very difficult.

If you can use AI to make people go much faster, or open it up so that many more developers can do it, that's when I think you get this explosion.

So to recap the three: it's a better brain, it's cheaper, and I think we've got AI coding tools now that can really help us go faster.

That's what a lot of Peppy is focused on, the intersection of these things happening to enable rapid robotics development.

Jared S. Taylor (01:27)

Most of the money in robotics right now is going into hardware and into models.

You picked the layer in between. What did you see that made you think that's where the bottleneck is?

Dan Lambert (01:40)

What I saw was my own experience of trying to set up a robot.

It took me over 10 hours to use the existing robot operating system, which is called ROS 2, Robot Operating System 2.

Very painful.

I have a lot of years of programming experience, and it should not be that hard.

You had to get the exact version of Ubuntu right. Just all the way around, it's not easy for people to enter.

My goal with this company was to take that initial onboarding of what used to be a 10-hour process and turn that into 10 minutes, where you are just getting started right away without having to deal with all of the configuration and setup underneath.

Let a platform handle that for you so you can focus on higher-level or higher-yield programming in very short order.

Jared S. Taylor (02:35)

You mentioned earlier how the cost of robots continues getting less and less.

Do you envision for many of us out there that we'll have a new robot every few years, just like we're trading in our iPhone or our Android phones?

Dan Lambert (02:51)

Yeah, I think that's kind of the model or approach of where it's going to be headed.

There's so much progress in a short amount of time.

A lot of that is that the cost of the actuators, which are the joints for the robot, has come down quite a bit. And they're becoming stronger.

So over time robots will be able to lift more load, even the bipedal robots.

You'll kind of trade in the robot.

Your robot works 24/7 for a few years, and it's got some parts that aren't working quite right, or it's doing the best it can with some failed components, probably kind of like my early Android phone.

You swap that out for a better model, and you kind of scrap the old one, a lot like we do with older phones.

But if we learn anything from the Android operating system, it's that the applications and the software and that whole developer ecosystem persisted throughout all the different versions of the phone.

That's actually the similarity we're trying to introduce in the industry now.

Even though you will have different robots over time, there will still be that single point of developer tooling, that single point of prompting, that single point where there are applications available to be loaded on the robot.

Easier said than done.

Jared S. Taylor (04:18)

What would you say is genuinely preventing general robots today?

Is it the models, the data, the dexterity, the reliability, the cost, the tooling?

What are kind of the main drivers preventing general-use robotics today?

Dan Lambert (04:33)

There are a lot.

We know solutions are coming, but there are some known limitations.

One of the big ones is dexterity.

And I would say dexterity at a reasonable price point, because there are hands out there that are very dexterous and very capable, but they're still at that $20,000 price point for just a single hand to get that level of manipulation.

So dexterity, combined with a cost reduction where you need to get down to a $2,000 or $3,000 hand that can do a lot, I think that's a big one.

We're also facing a major shortage of physical data in the world.

When the large language models were trained, they had plenty of data on the internet.

But nobody's really recorded a lot of sessions of people doing things in the real world, or what hand motions are being made, or things like that.

The founder of Figure has been building a data repository with a lot of this real-world data.

But I think we're going to see that from a lot of sources and a lot of companies that specialize in capturing data in certain scenarios, like a lab or a warehouse or whatnot.

The more of that kind of data that we get, I think the easier this will be.

If somebody requests something and there's no physical data, we use simple reinforcement learning to try to get an approximation of the thing that's being done.

But of course that's nowhere near as good as actual physically captured data that you can plug in and use as a base training set.

Jared S. Taylor (06:09)

Recently, Dan, you raised a pre-seed. Talk me through that a little bit.

This is not your first time raising capital. You have been through this time and time again.

But you and I did chat a few weeks back, and you mentioned something to me about the fundraising process, how cool it was to have so many people who were so interested in what you were building in this space.

Talk me through that a little bit, where you see the use of proceeds going with this pre-seed, what the team looks like, any details you can share.

Dan Lambert (06:46)

Absolutely.

I do think this is probably the most fun business I've worked on, just because everyone's excited about what that world of science fiction was 20 years ago now becoming real.

Almost every meeting I requested, people wanted to talk about robots.

And people wanted to talk about things they could solve in their immediate daily life, which was also very exciting to think about.

Something as simple as, "I want to water the plant," or, "I need to let my dog out in the middle of the day."

That was a common use case that we heard a couple of times.

Hearing all of that and hearing the excitement around it translated into a lot of venture investor interest.

We raised about a million. It was really oversubscribed.

We had considerably more interest than what we ended up going with.

I ended up going with VCs that I've worked with before.

You know how they operate. They promised to keep things pretty independent, which is always really important to me, to just be able to take the business where I think it needs to go.

We raised from Element Ventures, Rock Creek Capital and Mana Ventures, and then also from Colin Gardiner, who's a VC at Yonder who specializes in marketplaces.

That may not seem like an immediate fit, but this eventually will be kind of a marketplace for different apps and robot bodies and all of that, just a lot like how Android was.

We really wanted a marketplace expert on board, and we got that.

So we got to be choosy in who we brought on board.

And plenty of strategic angels that are involved in robotics wanted to jump in too.

Jared S. Taylor (08:27)

You mentioned when you were raising you'd have some investors that would give you use cases, right?

Like, "I want to be able to let the dog out a few times a day," or whatever.

What was the most complex use case they were hoping to be able to do?

Dan Lambert (08:43)

This one actually came from a lab, not an investor.

They wanted to be able to completely unpack this strange packaging, unscrew a bottle, and dip a really fine stick into the urine to test it.

It was a whole bunch of tasks in sequence that were pushing the edge of what a $5,000 robot can do right now.

We're really trying to stick to that cheaper form factor at the moment, just like that early Android phone.

It wasn't a super high-end phone. It was kind of janky.

That's the reality of the robots we're working with right now. They can only do so much.

It was this series of five tasks that were all really hard for it to do.

So then we just had to figure out, okay, how do we explain that we can only help you build a piece of this right now?

But give the industry a few years and it will have much more capability than it does right now.

Jared S. Taylor (09:41)

You said give the industry a few years.

It seems like when we've been estimating how quickly technology across the board, with these large language models, will move, there have been instances where we've heard AGI for a long time.

We've heard, "In six months, this role will be eliminated."

You mentioned a few years out.

Do you foresee it actually being a lot less than that?

And if so, what would need to happen to speed up that timeline?

Dan Lambert (10:12)

It's a complex question, but a good one.

Why I think it's a few years is simply because we need that time for data capture that we just don't have in the real world.

And we need to make mistakes in the real world too.

Part of this is trial and error.

The large language models got much better once they had a lot of usage, iterating on it and seeing better ways to do things.

Now it's a veritable arms race.

But with robots, we don't have enough experimentation in real-world conditions yet because of the limited number of robots that are actively in the field.

We estimate about 50,000 humanoids around.

But that number needs to be in the millions in order for there to be real progress of things running into walls and dropping an object, or using too much force or too little force.

We call this the sim-to-real gap.

You can now spin up simulations of a robot walking over rocks a thousand times, but then it gets out into the real world and steps on that one jagged rock that you didn't plan for, and all of a sudden it breaks or you have a very negative outcome.

Why I think it's a few years to have high-utility applications is because we need to deploy more robots and get punched in the face.

We need more users solving their daily problems or their daily business problems with these robots so that we collect enough data to understand how we fix these nuances that happen when you put robots into a world that they frankly don't understand or comprehend right now.

Jared S. Taylor (11:57)

Talk me through basically prompt to motor.

Just a random example, right?

I type, "Pick up that red cup and put it in the sink."

Walk me through every layer until voltage reaches a motor.

Dan Lambert (12:11)

Great question.

The first thing to understand is, why can't AI actually program these robots?

People are certainly trying to use AI to do this.

It's because, in large part, if you use an AI and that AI hallucinates or has a bad example, the robot can overextend itself, hurt itself or hurt a person.

There are these issues where you can't afford to allow AI to just really overrun your robot.

So we had to introduce a framework that we call PeppyOS, which is a layer where no matter what the AI builds or whatever the AI spins up, it can't damage the physical underlying robot.

It has to follow certain constraints that we can put in, whether that's harm to itself or others.

We spent the last year with four different engineers building that layer to make sure that AI can actually run on top of it.

To answer the question more directly, it starts with that prompt at a very high level, which is in English.

Then we break it down into a couple of concepts.

One is: what is the environment that the robot is going to be operating in?

Do you have a picture of that environment, or do you want us to just select a generic kitchen or warehouse or whatever?

Because the environment and conditions are just as important as what the robot itself will do.

Then, in a separate stream, it's setting up the robot itself in a simulation.

It goes basically from high-level instruction to setting up the environment and the robot via Python using the constrained framework that we have.

Then it runs a series of checks to make sure that nothing problematic will happen.

Then that translates it down into a series of Rust and Python commands, which are then translated into actual actuator movement.

I'll include a video that shows all the individual joints moving and exactly how much electricity is applied, et cetera.

But all of that is handled for the user.

That's behind the scenes. We control every single little joint in the robot and the voltage.

We're built on top of something called OpenArm.

In order for our platform to work, it has to work on open-source hardware because we have to be able to control every single joint.

We abstract those away basically as Python nodes that can be read.

So it's a little bit technical, but from a user standpoint, all they have to do is really prompt and then see what the robot does.

The benefit of our platform is that you can run it in simulation and also in the real world.

And that same code can be used in simulation and the real world.

We had to match up our simulation with what the robot's doing in the real world.

We use something called Isaac Sim, which is produced by NVIDIA, to actually do a lot of the renderings of these robots in actual motion.

Huge help.

NVIDIA has been amazing at developing a lot of tools for us to use in this process.

Jared S. Taylor (15:32)

What would the state of robotics look like if NVIDIA didn't exist?

Dan Lambert (15:37)

We wouldn't have a robot brain, and we wouldn't have a good simulator.

There are other robotic simulators, but they really just aren't as good.

They run on a CPU instead of a GPU, so the performance is significantly degraded.

Also just generally, NVIDIA makes amazing GPUs that allow this real-time interaction and responsiveness.

I think the industry would be set back several years, or somebody would eventually need to develop that level of chipset for this robot revolution to actually happen.

Jared S. Taylor (16:11)

I don't want to picture a world in which NVIDIA doesn't exist, but I just wanted to get your take on that.

Dan Lambert (16:17)

Yeah. We're big fans here, obviously.

We use a lot of what they develop.

Jared S. Taylor (16:23)

Talk me through who you would say are the clear-cut leaders.

Who is falling a close second and third in your mind?

Dan Lambert (16:38)

I think it depends on which piece of the market you're going after.

In terms of full vertical integration, where you own the robot and the environment and you're developing robots for a specific solution, I would say Tesla Optimus is probably number one.

It's going to be very pricey to start. They're using a lot of very expensive components.

I think a lot of the world doesn't need a $70,000 or $80,000 robot.

I think you need the $5,000 version.

But in terms of capabilities and quality of manufacturing, the Tesla units are really up there.

We're definitely seeing Figure as a solid competitor, also vertically integrated.

The company that's sold kind of the most units is based out of China, so I think we're going to see some more of those.

Then on the consumer side of the market, we're seeing a lot of progress from some toy robots that have gotten more advanced.

There's the SO-ARM101.

It's just a little mechanical arm for development, and we've seen tons of those units sold for developers who are trying to enter the market for the first time.

Jared S. Taylor (17:59)

You and I recently caught up, but we caught up one other time before that.

I think it was around last year or the year before.

You were just kind of dabbling and building some robots, I think, in your garage or in your house.

Talk us through what you've learned since going from dabbling to making this a full-time venture.

Dan Lambert (18:26)

We started exploring this early, and I think we needed to understand it by trying to solve a very specific use case ourselves.

We started with trying to pack boxes.

That was just for learning. It wasn't necessarily for a particular client or anything.

But in going through that process, you feel all the pain points of where the industry is right now.

How hard it is to know which robot to use.

How hard it is to get the right environment set up.

How to deal with physical limitations.

How to configure the robot brain to be responsive.

In a packing environment, things change.

How do you handle a truly difficult environment for robots?

As we did that as a practice run, getting to a packing robot, we realized this would have been way faster if we had been able to use AI to prompt through a lot of these sequences using all of the existing industry components.

Having gone through that experience, we realized there's definitely a business here in AI infrastructure and developer tools.

We can make this process way better for your average Python developer, or somebody who's been building agents or using tools like Clawbot.

We're trying to attract that semi-technical user who can actually start enabling their robot.

I think that's when this market gets super exciting.

It's when all these millions of people who are semi-technical really want to start playing with things and solving problems.

Our goal is to open that aperture and enable millions of people to start iterating on robots.

There's an exciting moment of: what are people going to come up with?

What are they going to build when you enable the masses, or even just semi-technical people, to start building robots?

That's the moment we want to capture.

Jared S. Taylor (20:41)

This is something we've been seeing more with the coding platforms, right?

Claude Code and Replit and Lovable, where we have these semi-technical people starting to build where, in the past, they would only be able to build if they had a software engineer they were relaying their vision to.

I know you've used that tagline, "Lovable for Robotics," in some of your framing.

What does that style of building have to survive now that it never had to survive before?

Dan Lambert (21:16)

That's a great question.

Even though the tagline is "Lovable for Robotics," the reality is that it's actually a drastically different problem.

It's one thing to build a website on the fly, and it's completely different to have a robot in your house all of a sudden scrambling eggs.

Those are very different problems.

The one thing that is common, though, is that both solutions are designed to be rapid prototyping.

That's what we're going for.

Can you get your robot 90% of the way there in simulation, knowing that before you have to buy an expensive piece of hardware, you know exactly what the limitations are and what it can and can't do?

We generate that open code.

You can take the code that our system generates, take it to a robotics engineer, and maybe they solve that last 10%.

Very similar to how Lovable, a lot of times, doesn't take you all the way to the end of the application.

You might need security, or you might need some advanced feature or something like that.

That's not the purpose.

The purpose isn't a final, finished application.

The goal is to get you 90% of the way there, make sure this is a viable project, then you buy the robot, and then you can finish out the remainder of the development to solve your problem.

One of the biggest differences is just how many sheer hours of simulation you're going to have to run with the robot.

With a website, it's just a website. It works on all browsers the way they've set it up.

The physicality of interfacing with the real world means you have to run simulations over and over again and think through the universe of variables of what could go wrong.

You have to really think through your environment and what happens when something dynamically changes.

What happens when there's a dog in the way?

These are things that a website development platform just doesn't have to think about.

But those are realities that we have to deal with.

Dan Lambert (23:06)

There will be some amount of journey of us figuring that out as we release this platform.

We're releasing Peppy to the public so anybody can access the web prompt directly on the homepage starting in December 2026.

We're excited to see what people build.

We've been sending out some surveys about what kinds of problems people are trying to solve, and it's been really exciting to see people's imaginations and what they want to do with it when it's available.

Jared S. Taylor (23:50)

Super exciting.

That launch will be here before you know it.

Dan Lambert (23:54)

Yeah, that's coming quick.

Jared S. Taylor (23:57)

How will you and your team like to receive feedback?

Say you go live December 2026.

What's the best way you like to receive feedback, and in what format, so it can actually be used to have an action?

Dan Lambert (24:18)

I think we'll see all of the user prompts.

We'll be able to see where they get stuck, what didn't work, and also what kind of underlying robot they're trying to export down to.

We're also going to have a customer service team that people can call and explain what they're trying to do.

We'll figure out some things in the background.

The reality is it's not going to be a finished product day one.

There's a lot of iteration, and we really want to hear from users what they want to solve and also where they got stuck.

If there's a feature they really need but we simply don't have it yet, part of what we need to do as we get these requests is go out to the market and say, look, we need this scrambled eggs app skill.

We're just not great at it right now. We've got an approximation, but there are a lot of errors in it.

We really need someone who's been building this kind of thing.

Plenty of companies have been doing the folding-a-shirt app. We need that. Or handling laundry.

There's such a big universe.

What we're seeing is that the industry has a ton of point solutions.

If we know what users need, we can start partnering with all of these developed applications to bring them onto our platform.

That's part of the vision here.

Jared S. Taylor (25:47)

Do you foresee having a lot of similar solutions being prompted?

Dan Lambert (25:53)

So far, yes.

We've seen some regular patterns of things that people are trying to solve.

But then I would say 50% are not common.

They're kind of way-out-there requests.

It's like, "Hey, I need a bartender."

Some very abstract task where I'm like, I didn't think of that. Nobody else prompted that.

But they want it in their house making drinks when they have guests over or whatever.

There were some more explicit details on that particular request as well, but we don't need to go into that.

When you're opening up the aperture to the world of what people can prompt, you get some pretty odd requests.

Jared S. Taylor (26:38)

Are there some that you're going to push back on?

For instance, from a safety perspective, if someone wanted to have a barber with a straight edge and stuff like that.

I'm not even saying someone's trying to be malicious, but there may be dangerous prompts going in there.

How do you foresee dealing with those?

Dan Lambert (27:04)

That's a very good question.

As a policy, we're not allowing the robots to wield any kind of sharp object.

We're actively blocking a robot accessing a gun or a knife or anything.

But even that has shades of gray that we haven't really gotten to yet.

Let's say the robot wants to butter toast.

Well, guess what? The best mechanism to do that is to grab a butter knife.

Should we allow that because it's a dull knife?

But what happens when the robot grabs a sharp knife instead because that's all it can do?

That happens to me daily. I'm out of butter knives, so I just grab something else.

There's some gray space.

I do think we'll allow users latitude, but this also means a very big disclaimer on the site saying, look, you still have to use discretion in how you're telling the robot what to do and what to pick up.

We might not be able to distinguish between something that's very harmful and something that's just fine for the robot to pick up.

That's a reality of the early days of exploration.

But it's kind of an exciting problem to think about.

What do we enable and what do we not?

We don't have all the answers yet.

Jared S. Taylor (28:30)

I'm really interested in seeing what robots can do differently than how we as humans do it as well.

I'll give you an example.

There's some company out in Italy that's working on a robot to clean the clay tennis courts after you play, because you have to sweep them and then dust off the lines.

In my mind, it's like you need one of those iRobot-style things that kind of goes back and forth, like the lawnmower types or a Roomba.

They had a different approach, though.

They're like, because you're using a robot, you don't need to do it the way a human would do it.

They had something where it's one whole line that goes up and back to clean the court in one swoop, versus going back and forth and back and forth.

I've really started thinking about that.

You mentioned the knife thing.

If I don't have a knife, I go to a spoon and I warm up the butter and then I use the back of the spoon.

I'll be very curious to see how we're using robots to execute tasks that humans would usually execute, but in a much different, and maybe much more efficient, way as well.

Dan Lambert (29:49)

We're kind of leaning toward the world being a little bit more humanoid, mostly because a lot of the tasks that were described can at least be done with a humanoid.

Which makes sense.

Our world was built for people to do things.

Since so much of our existing infrastructure has enabled humans to do things, we know the human form factor will work in most cases.

But it won't be as efficient as a car or whatever.

It's like the robot Olympics.

There's a robot faster than any human on Earth, but a car is still better.

A car can go much faster than a bipedal robot.

So yeah, you can solve these things with a humanoid, but should you?

In a lot of cases, no. You can find a better form factor.

I do think, though, once people start developing applications on our platform, or at least contributing skills on our platform, as there becomes more of a corpus of apps and data and humanoid data, there's probably some convergence.

Why would you go with a different form factor?

Just deal with the inefficiencies of being human, because that's what we've always done, just dealt with the inefficiencies of being human, knowing there's something lost there in the process.

Jared S. Taylor (31:15)

Super interesting.

I'm really excited to see how your platform continues to evolve. I know you are as well.

Obviously that big launch and getting that feedback moving forward from there will be super important.

Dan, where did your love for robots come from?

When did it happen?

Everyone has that moment, right?

Everyone I've met in robotics has that moment in their life, and then they weren't just interested in robots.

They usually became obsessed with robots.

When was that moment for you?

Dan Lambert (31:45)

My undergrad is in computer engineering.

One of your very first projects in computer engineering is that you have to design a vending machine from scratch.

You're learning the concept of state machines.

You put in a quarter, and then the machine stays in a state until it's ready for the next action or whatnot.

I just fell in love with that kind of problem solving.

Wow, how do we make automation so humans don't have to do these things?

I spent all night on it.

Even though just the first part of the task was due, I was so excited to solve the whole vending machine.

I think that's when the first, "Wow, I love solving this kind of system problem," kicked off.

Of course, at the time, this was over 15 years ago.

The hardware wasn't even close. We had no AI.

A humanoid robot was a pipe dream at the time and would have cost like $300,000.

But I think the more recent magic moment was when I was able to just type in, "I want the robot to wave."

Without having to troubleshoot a bunch of stuff along the way, the robot gave me this half-hearted, "Hey."

The feeling was, man, this is a glimpse of the future where at some point you'll be able to just talk to the robot and it'll do it.

Even though that's 10 years away, you can feel that magic moment.

We're going to have a lot of help.

There are going to be brains that are bigger than ours.

There are going to be robots that are more capable than we are.

As a species, to be building something that exceeds our limits as humanity, what an exciting moment to be alive.

Jared S. Taylor (33:44)

What do you want to see happen over the next year in the robotics field that will continue to help us move forward in this area, in this industry?

Dan Lambert (33:54)

I realize we're on the software side, so we're not helping as much on this, but we need a ton of companies making a ton of robots.

As people start to experiment and simulate and try things, we need those new entrants developing a $2,000 to $6,000 robot that can start doing things and that everybody can start learning from.

I would say we're just in the industry phase of early learning.

We need more hardware to be able to do that.

Jared S. Taylor (34:28)

Dan, always a pleasure catching up with you and hearing what you're working on.

Super excited for Peppy, and let's definitely stay in touch.

Look forward to hearing more.

Dan Lambert (34:39)

Great. Thanks, Jared. Talk soon.