Reproducing DINO-WM

This tutorial illustrates how stable-worldmodel can be used to build an end-to-end world modeling pipeline. In the example below, we will produce all the code necessary to train and evaluate the DINO World Model (DINO-WM) on the PushT environment.

More specifically, we will cover:

1. Set up the world environment (PushT).

2. Collect a dataset of environment interactions from a random policy.

3. Pre-train DINO-WM predictor on the collected dataset using the stable-pretraining library.

4. Leverage the trained model to perform planning with Model Predictive Control (MPC).

5. Evaluate the performance of the trained model in the world environment.

6. Visualize the agent’s behavior.

---

Note

💡 Fun fact: the original DINO-WM repo had 14,903 lines of code. We managed to recreate everything in less than X lines!