nicklashansen/dreamer4
Unofficial implementation of the Dreamer 4 world model in PyTorch.
What it solves
This project provides an unofficial PyTorch implementation of the Dreamer 4 world model, allowing researchers to experiment with and extend a scalable architecture for predicting environment dynamics. While the original paper focused on discrete actions in Minecraft, this implementation adapts the model for continuous control across multiple tasks.
How it works
Dreamer 4 utilizes a block-causal transformer architecture split into two main components:
- Causal Tokenizer: Encodes image patches and latent tokens, compressing them through a low-dimensional projection to allow frames to be decoded sequentially.
- Interactive Dynamics Model: Operates on a sequence of actions, noise levels, step sizes, and tokenizer representations, using a shortcut forcing objective to denoise representations.
Who it’s for
It is designed for AI researchers and developers working on world models, reinforcement learning, and continuous control tasks, specifically those who prefer PyTorch over JAX.
Highlights
- Multi-task Support: Trained on 30 continuous control tasks from DMControl and MMBench.
- Comprehensive Dataset: Includes 7,200 trajectories (3.6M frames) of expert and mixed-quality data.
- Interactive Interface: Comes with a web-based UI to interact with the trained world model in real-time.
- Ready-to-use Checkpoints: Provides pre-trained tokenizer and dynamics model weights via HuggingFace.
Related
- Project
- Project
- Project
- Project
- Project