Bot Eat Brain
Posts
DrEureka trains robot dog to balance on a ball without real-world training

DrEureka trains robot dog to balance on a ball without real-world training

PLUS: The news just got Grokked

Michael Parrish
May 06, 2024

“Get in, quick! There’s no time to explain!”

Who do you trust? 👇

Choose wisely. Your life depends on it.

Good choice, that was close. 😰

Ideas? Comments? Complaints?

Respond to this email or hit me up on 𝕏.

Now on to the newsletter. 🤗

TOGETHER WITH

Good morning, human brains, and welcome back to your daily munch of AI news.

Here’s what’s on the menu today:

Robot dog balances on a yoga ball 🐶 🧠
DrEureka trains robots without any real-world training or fine-tuning.
Make 3D models and scenes in under a second 🧊 🏜️
GS-LRM generates high-quality 3D scenes from photos.
The news just got Grokked. 📲 🗞️
𝕏 launched AI summaries powered by Grok.

Sponsor Bot Eat Brain (Now 50% off) 🤯

New here? Subscribe! 😎

MAIN COURSE

Luke, I am your paw-ther 🐶 🧠

On Friday, NVIDIA, UPenn, and UT Austin unveiled DrEureka.

It automates the entire sim-to-real process for robots. It allows them to learn new skills in simulated environments and deploy them in the real world.

What does this look like?

With this method, they trained a robot dog to balance/walk on a yoga ball without any real-world training or fine-tuning required.

Source: DrEureka

That dog is… on the ball. Amirite? 😏

How does it work?

It builds on NVIDIA’s previous work with Eureka, an algorithm that taught a five-finger robot hand to spin a pen.

For more on Eureka, click here.

DrEureka is an AI agent that writes code to train robot skills in simulations. It then writes more code to fully automate the sim-to-real process.

Robot dog? Yoga ball?

The ball’s bouncy surface is difficult to accurately simulate. Despite this, DrEureka can balance, steer on multiple terrains, walk sideways, and even stabilize itself after the ball is kicked.

Source: DrEureka

This whole process is automated?

Correct. Sim-to-real transfer typically requires expert roboticists to adjust every parameter by hand. DrEureka leverages AI models like GPT-4 to account for gravity, friction, and more. It automatically tunes for all parameters and articulates its reasoning accurately.

That’s dog-gone amazing.

Paw-sitively. 🐾 It’s open-source on GitHub, so you can take the codebase out for a spin. If that’s your thing, click here.

SIDE SALAD

Instantly create 3D models & scenes 🧊 🏜️

Last week, Adobe and Cornell University introduced GS-LRM.

It generates high-quality 3D scenes from 2-4 photos taken from different angles. It does this in a fraction of a second.

Source: GS-LRM

GS-LRM?

It’s an AI model that breaks photos into small pieces, processes them, and reassembles them into detailed 3D models or scenes.

What’s so great about it?

It outperforms state-of-the-art methods with its straightforward Transformer architecture. It approximates the fine details and overall shapes of its 3D reconstructions with high computational efficiency.