Mels is an AI robotics sandbox. Build a robot with Lego-like blocks, train it with AI, and watch it take on challenges — no code required.

Get Started 👉 mels.ai to make your first Mel in minutes!

• 🌐 Fully browser-based — access instantly online: no downloads, no installs, no local GPUs

• 🎯 Custom RL environments — design new behaviors with custom rewards and training objectives

• 🧱 Modular morphologies — build any body from blocks and joints: humanoids, quadrupeds, wheeled, hybrids

• ⚡ Real-time training — simulate millions of steps in minutes, live in your browser

• 🆓 Free on-device inference — train once in the cloud, interact forever locally

• 🤝 Creative community — share and discover creations on Discord

Learn More 🔗 mels.ai · Mels Dev Discord · Mels on Substack · @mels_ai on X · @mels_ai on YouTube

FAQ

What is Mels?

Mels is a creative platform for a new category of AI (a “Mel”) that’s neither an “agent” nor a “robot” but lives at the intersection of the two:

• On one hand, Mels are digitally native and powered by state-of-the-art AI, including deep neural networks and reinforcement learning for sequential decision-making and actions

• On the other hand, Mels are physically intelligent and embodied like robots, with the ability to sense, learn, and act based on real-world physics simulations

You can think of them as self-learning digital robots. But we like to call them Mels.

The Mels core loop is:

• You see a challenge, like a creative prompt or a competitive leaderboard

• You build a Mel with Lego-like blocks and joints, then train it with AI

• You watch it learn on its own from scratch and take on the challenge

• You keep tinkering to turn epic fails into shareable wins

• You unlock new building blocks, creative tools, and challenges

All from your browser. No coding, AI, or robotics expertise required. With a community of makers like you.

How does Mels work?

Mels is powered by embodied AI systems invented at Google Research, combining massively parallel real-world physics simulation, state-of-the-art deep reinforcement learning training, and efficient model inference to run complex robot AIs on any consumer device.

Traditionally, robotics simulation has required a lot of money, time, and expertise just to get set up. Then the experiments themselves take hours or days, leading to slow feedback loops and iteration. And finally, sharing experiments or reproducing others’ work is painfully clunky.

Now, one tap = instantly access, train, and share AIs with anyone, anywhere, on any device.

Mels turns your browser tab into an AI robotics sandbox where you can:

• Train custom embodied AI agents in robotics simulation

• Watch them learn and interact with them in real time

• Share and load experiments instantly via a link

No cost. No setup. No barriers.

Who is Mels for?

Mels is for anyone who wants to create their own AI robot — no coding, AI, or robotics expertise needed. This includes:

• Consumers playing in a creative sandbox

• Students learning and educators teaching AI concepts hands-on

• Creative technologists exploring a new interactive medium

• Hobbyists and tinkerers who love making things

• Game developers designing AI characters and behaviors

• Researchers experimenting with deep reinforcement learning and robotics simulation

If you’re curious about AI and want to train your own robot, Mels is for you.

Why does Mels exist?

So much about AI is undefined, but one thing we know for sure is it’s changing everything. That feels big and scary but it also means we’re all in it together. We’re a generation defined not by age but by our shared experience of living through AI’s impacts.

There are many ways people, companies, and institutions can choose to exploit AI. It often starts with automating existing things to save time or money. And then maybe creating new AIs for you to use.

But you know what’s cooler than using AIs others make? Creating your own.

Creating your own AIs may sound like science fiction, intimidating, or just plain boring…but it doesn’t have to be! One of the special things about AI is that we can now make the experience of creating AIs as joyful and accessible as play. One everyone can take part in.

We believe creative play is the best way to put the power of creating AIs into everyone’s hands. Because time and again, people will surprise you when you give them the power to bring their imaginative creations to life.

(Spoiler alert: extraordinary things happen :)

About Luda

Luda is a new kind of AI lab that believes **play is the highest form of research.**

We’re a team of veteran misfits who fuse deep technology (Google Research, SAP, Magic Leap) with interactive media (Disney, Anki, Sims). Our investors and advisors include Jeff Dean (Chief Scientist at Google), Illia Polosukhin (co-author of “Attention Is All You Need”), Bing Gordon (ex-CCO of Electronic Arts), Scott Belsky (Partner at A24), and senior researchers at Google DeepMind and Anthropic.

🔗 ludaresearch.com · @ludaresearch on X · Luda on LinkedIn

At Mels, we're building a robotics simulation sandbox where play comes first. Because when you remove barriers, you unlock breakthroughs.

Think about it:

• 🔲 Children learn complex physics by playing with the world around them (like stacking and toppling building blocks)
  

• 🎮 Gamers solve problems researchers thought impossible (like when Foldit players solved an HIV protein structure that stumped scientists for 15 years)
  

• 🧑‍🔬 Scientists discover breakthroughs by "playing" with ideas (just ask Fleming and his accidental discovery of penicillin)

It’s no coincidence, it’s human nature.

That's why we're making state of the art robotics as creative and fun as play. No complex tools or technical barriers. Just intuitive, accessible play.

And we’re in good company. We and our friends at Google DeepMind collaborated on research to bring robotics to life in your browser: research.mels.ai.

Build, train, and play with state of the art robot AIs. All in your browser. No setup. One tap away.

The next breakthroughs in robotics won't only come from closed labs. They'll come from anyone, anywhere exploring and discovering through play.

Because play isn't just fun - it's fundamental.

Imagine having an entire AI robotics lab in your pocket that runs right in your browser.

No fancy hardware, no large downloads, no complex setup. Just one tap and you're in.

Till now, robotics simulation experiments have required a lot of money, time, and expertise just to get started.

Then the experiments themselves - the simulation, training, and inference - take days or weeks, leading to slow feedback loops and iteration.

And even after all this, sharing your experiments (the code, data, and models) or reproducing others’ work is painfully clunky.

With Mels, your browser tab becomes an AI lab where you can:

• Train custom embodied AI agents

• Watch them learn in real-time (seconds, not days!)

• Share them instantly via a link

The unlock: our AI physics engine combines efficient model inference with powerful real-world physics simulation to run complex embodied AIs in on any consumer device.

As a result, every Mel is an AI that’s:

• Physical: understands real-world physics

• Immersive: rendered in 3D simulations

• Interactive: perceives and acts in milliseconds

All in your browser.

One tap = instantly access, train, and share AIs with anyone, anywhere, on any device.

No setup. No wait. No barriers.

The future of AI isn't our research lab — it's yours.

Starting From Scratch

Every Mel begins as a blank slate—no prior experience, no idea how to move—like a newborn, but without any evolutionary instincts. Since Mels are uniquely built from Lego-like blocks, with entirely different morphologies, each one is kind of a new species! 

To learn how to move, Mels must start from the ground up, mastering every step through trial and error in a real-world physics simulation we like to call the Meltiverse.

Enter The Meltiverse

The real fun begins when your Mel is ready to train. Our system dynamically assembles a training simulation from the unique body, rewards, and environment you give it:

It then generates thousands of Mel variants, all training simultaneously in custom-simulated worlds:

Imagine thousands of your Mel, experimenting with different movements, receiving feedback on what works and what doesn’t, and sharing their findings with one another. It’s a large-scale, collaborative learning process fueled by the unique experience of each individual Mel. Your Mel then taps into this collective experience to learn to move better and better with every training update.

The Training Process

The typical training process involves simulating and learning from tens to hundreds of millions of steps. Traditionally this used to take hours or days but now it takes seconds and minutes. It’s only possible due to a combination of creative design choices, novel algorithmic advances, and massively parallelized, hardware-accelerated simulation.

So now, thanks to the Meltiverse, everyone can train digital robots to life.

Sports have always been a big source of inspiration for Mels, from specialized training to physical play to entertaining performances. And of course, all the epic wins…and fails.

Now, any matchup of Mels can race down the track, test their physical abilities, and create emergent chaos. It's going to be weird, fast (maybe), and funny (definitely).

Let the games begin!

This Mel is trained using Deep Reinforcement Learning, a technique where AI agents learn to accomplish a goal based on lots of trial and error as well as being rewarded when they do something correctly. Every new combination of blocks learns to move in its own way based on custom training rewards and real-world physics simulation.

Mels are physically embodied AIs that represent a whole new kind of media that's only now possible thanks to state of the art technology.

Endless possibilities, ridiculous fun.

Let’s dive in!

The future of AI is embodied.

To realize the promise of AI, we believe it needs to be grounded in physical understanding and interaction. But this isn't as easy as it sounds.

Physical action may seem straightforward compared to job automation or artistic expression, but for AI, it's exactly the opposite. What we perceive to be hard as humans is actually easy for computers, whereas what we perceive to be easy is actually hard.

There’s an obscure contradiction known as Moravec’s Paradox that summarizes this perfectly:

It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old regarding perception and mobility.

But why?

Moravec continues:

Encoded in the large, highly evolved sensory and motor portions of the human brain is a billion years of experience about the nature of the world and how to survive in it. We are all prodigious olympians in perceptual and motor areas…so good that we make the difficult look easy. Abstract thought, though, is a new trick, perhaps less than 100 thousand years old. We have not yet mastered it. It is not all that intrinsically difficult; it just seems so when we do it.

As a species, humans have spent a shockingly small fraction of our evolutionary time learning abstract thought compared to physical motion. So it's unsurprising that what seems simple to us, is so difficult for AI.

If the future of AI needs embodiment, we need to brave Moravec's paradox to achieve it.

That's why we're developing Mels.

Mels are digital robots that are physically simulated and powered by state of the art AI, including deep neural networks and reinforcement learning for sequential decision-making and actions:

They're physically intelligent and embodied like robots, with the ability to sense, learn, and act based on real-world physics simulations. They aren't just computing - they are embodying:

While many are crowding into applying the current state of AI, we're veering off the beaten path to help bring embodiment to AI's future.

Two roads to AGI diverged in a wood, we took the one less traveled.

Then we made it possible for anyone to build a Mel:

Now we’re making it possible for anyone to teach their Mel new things, starting with custom training rewards!

By default, Mels are trained using a universal reward function in our general-purpose training and simulation system. This is so anyone can build and train a Mel to learn to move from scratch without having to be an AI researcher or roboticist to get started.

Custom rewards is a first step in opening up our AI machinery and giving people even more creative control over what and how their unique Mels learn - all while keeping things playful and accessible.

We’re in uncomfortably uncharted territory for consumer AI. But one thing’s for sure: things are going to get weird, and we’re here for it.

Watch each Mel strut down the catwalk and show off its one-of-a-kind rizz!

Mels are a new kind of digital robot where every new combination of blocks learns to move in its own way based on real-world physics.

We’ve said it before and we’ll say it once more: time and again, when you give people the power to bring their imaginative creations to life, extraordinary things happen.

We're early, but if you want to build your own Mel, sign up for playtesting.

We coined “Mels” to name a new kind of AI that's neither an “agent” nor a “robot” but lives at the playful intersection of the two:

• On one hand, Mels are digitally native and powered by state of the art generative AI, including deep neural networks and reinforcement learning for sequential decision-making and actions

• On the other hand, Mels are physically intelligent and embodied like robots, with the ability to sense, learn, and act based on real-world physics simulations

So you could think of them as digital robots. But we like to call them Mels.

That said, Mels are unlike both agents and robots in one crucial way: you don’t just use Mels…you create them!

That means you can build and train Mels with a wide range of forms and functions.

It also means Mels learn how to use both their minds and bodies, starting from a blank slate.

But the best part is: anyone can make Mels, without being a physicist, AI researcher, or roboticist.

Training is fast, inference is free, and interaction is real-time - all from your browser, powered by machinery we developed ground-up for Mels:

We're early, but if you want to go deeper, sign up for playtesting.

Watch 10 totally different Mels get assembled from the same 5 basic types of Lego-like building blocks.

Mels are a new kind of physically intelligent AI where every new combination of blocks learns to move in its own way based on real-world physics.

Each Mel struts down the catwalk and shows off its one-of-a-kind rizz.

These Mels are trained using Deep Reinforcement Learning, an Artificial Intelligence technique where AI agents learn to accomplish a goal based on lots of trial and error as well as being rewarded when they do something correctly.

Mels are physically embodied AIs that represent a whole new kind of media that's only now possible thanks to state of the art technology.

They're also just ridiculous.

@ai_mels5 Parts 10 Mels (Deep Reinforcement Learning) In this video, 10 totally different Mels are assembled from the same 5 basic types of Lego-like building blocks. Mels are a new kind of physically intelligent AI where every new combination of blocks learns to move in its own way based on real-world physics. Each Mel struts down the catwalk and shows off its one-of-a-kind rizz. #reinforcementlearning #ai #neuralnetworks

Tiktok failed to load.

Enable 3rd party cookies or use another browser

Some talk, ours walk.

Let’s get into it…

AI is making a leap from “models” to “agents”.

Agents can sense, think, and act autonomously - and they can learn and evolve forever from their experiences with humans, other agents, and the world they live in!

Even in their primitive state, it’s not crazy to say modern AIs are almost like a new kind of lifeform.

Most AIs today are:

• conversational - think texting but with an AI

• 2D - can “hear and speak” language and “see and draw” images

• disembodied - have a name or avatar, but fundamentally is a "brain in a vat"

Our AIs are different:

• immersive - think playing but with an AI

• 3D - can sense, think, and act spatially in digital worlds

• embodied - physically simulated and controlled

The coolest thing about our kind of AIs is that you can build, train, and evolve them just by playing, without ever needing to engineer prompts, animation, or code.

They’re created brick by brick - by you.

If AIs are a fundamentally new medium, our user-generated AIs make it a *creative* medium.

And as we’ve said before, when you give people the superpower to bring their creations to life, they’ll surprise you.

In this video a Mel, which is an AI that's assembled with Lego-like building blocks, comes to life, takes its first steps, and learns how to move to a target.

Initially the target is behind a finish line on a flat (but still treacherous!) catwalk, but the legged Mel succeeds.

When the finish line is moved to the other side of an icy ramp, the legged Mel struggles and needs to mod its legs with wheels to reach the target.

This Mel is trained using Deep Reinforcement Learning, an Artificial Intelligence technique where AI agents learn to accomplish a goal based on lots of trial and error as well as being rewarded when they do something correctly.

Mels are physically embodied AIs that represent a whole new kind of media that's only now possible thanks to state of the art technology.

More Mels coming, subscribe to our new YouTube channel.

AI Generation

So much about AI is undefined, but one thing we know for sure is it’s changing everything. That feels big and scary but it also means we’re all in it together.

We’re a generation defined not by age but by our shared experience of living through AI’s impacts.

UGC

There are many ways people, companies, and institutions can choose to exploit AI. It often starts with automating existing things to save time or money. And then maybe creating new AIs for you to use.

But you know what’s cooler than using AIs others make? Creating your own.

Play

Creating your own AIs may sound like science fiction, intimidating, or just plain boring…but it doesn’t have to be!

One of the special things about AI is that we can now make the experience of creating AIs as joyful and accessible as play. One everyone can take part in.

Purpose

As a team dedicated to lifting people through play, UGC gameplay is the best way we know how to put the power of creating AIs into everyone’s hands.

Because time and again, people will surprise you when you give them the power to bring their imaginative creations to life.

(Spoiler alert: extraordinary things will happen :)

We started R&D on novel generative AI techniques five years ago in Google Research with the belief that agents will be a fundamentally new medium for AI. Just like cameras enabled photos, computers enabled video games, and the internet enabled social media, AI will enable agents.

But the agents we’re dreaming up are nothing like what people think agents are today. Ours are:

• 🤖 Interactive: perceiving and acting in milliseconds

• 🌐 Immersive: embodied and rendered in 3D

• 💪 Physical: simulated with real-world physics

In theory Reinforcement Learning is the ideal technique for these types of physically intelligent 3D agents. But in practice it's a slow, expensive, and offline process. Enter: Real-Time Reinforcement Learning to make our agents fast to train, cheap to run, and online. RT-RL unlocks:

• ⏱ Accelerated Learning: from hours and days of training time to seconds and minutes

• ⚡ Low-Latency Action: from offline inference on high-end servers to online inference in any consumer browser

• ➰ Human-In-The-Loop: from static 2D videos to interactive 3D agents

Today, RT-RL enables a new kind of agent for a new kind of media. Tomorrow, we make this a creative medium by empowering you to create agents in an experience that’s as joyful as play.

We like to say we’re building UGC gameplay for the AI generation. If that’s you, subscribe to stay tuned.

Subscribe now