Dreamer4 Implementation

Custom implementation of "Training Agents Inside of Scalable World Models" originally published by Google DeepMind. Trained on 8x 16GB V100 GPUs using data from zhwang4ai/OpenAI-Minecraft-Contractor.

Key Differences from original paper:

• Uses MSE loss only when training tokenizer (no dynamic LPIPS integration as of yet)
• Tokenizer/World Model/Imagination Training etc... all overfit on one video to prove the pipeline works on limited compute resources.
• Due to limited computing resources, tokenizer and world model outputs are grainy than ideal reconstructions.

Results

Tokenizer Performance

Example #1: Reconstruction from Tokenizer

reconstructed_output.mp4

Example #2: Reconstruction from Tokenizer

v2_reconstructed_output.mp4

World Model Performance

Single-step inference (shows that overfitting worked)

one_step_inference.mp4

Multi-step inference (simulates actual world model performance)

multistep_inference.mp4