Lecture 4 April 11, 2026

World Models

Teaching robots to imagine. Build IRIS, DINO-WM, and DIAMOND world models from scratch in Colab, then scale to real robots with LeWorld and DreamZero — the first open-sourced world-action model for SO-101.

IRIS DIAMOND DINO-WM DreamZero LeWorld Video Diffusion
Part 1: Foundations Part 2: Scaling to Real Robots

Practical Hands-on Notebooks

Build three world model architectures from scratch and train them on robot manipulation tasks.

📗

IRIS World Model

Build IRIS from scratch — VQ-VAE tokenization, autoregressive transformer, and imagination-based RL for Atari.

Open in Colab
📘

DINO World Model

Build DINO-WM from scratch — predict in DINO feature space instead of pixels for robust, generalizable world models.

Open in Colab
📙

DIAMOND World Model

Build DIAMOND from scratch — diffusion-based world models that generate crisp future frames via iterative denoising.

Open in Colab
📕

DINO-WM on PushT

Apply DINO-WM to the PushT robot manipulation task — train a world model and plan actions in feature space.

Open in Colab

Lecture Outline

Two parts covering the full world models landscape — from game-playing agents to real-robot deployment.

Part 1: Foundations

🌍

What Are World Models?

🎮

IRIS

🦕

DINO-WM

💎

DIAMOND

🎬

Video Tokenizers

🧩

DiT Architecture

Part 2: Scaling to Real Robots

🤖

LeWorld

💭

DreamZero

🦾

SO-101 Fine-Tuning

📊

Evaluation

🔮

Future Directions


Interactive Visualizers

Explore DreamZero's world-action model through an interactive gallery and research site.

Research Site
🌐

DreamZero-SO101

Full research site with interactive rollout gallery, paper, and evaluation results for our open-source world-action model.

Visit site
Interactive Gallery
🎥

Imagined Rollout Gallery

Pick a training example, choose a prompt, and watch DreamZero imagine the robot's future actions and video frames.

Try it