Mini Dreamer is an experimental interactive world model that learns to predict future observations while you explore an environment in real time.
The system continuously improves its predictions using online learning, replay buffers, and latent dynamics, while visualizing prediction quality and training metrics live.
This project serves as a minimal platform for exploring predictive representation learning and world models.
Click the image to watch the demo video.
The interface displays:
| Real Environment | Model Prediction |
|---|---|
| Ground truth frame | Model's predicted next frame |
Additional panels visualize:
- prediction error heatmap
- learning metrics
- training curves
- replay buffer statistics
- latent state information
As the agent explores the environment, the model gradually improves its predictions in real time.
Explore a procedurally generated environment while the model learns.
WASD movement
camera rotation
minimap visualization
The system learns a predictive latent representation of the environment.
Pipeline:
observation_t
↓
encoder
↓
latent state (z)
↓
RSSM recurrent dynamics
↓
decoder
↓
predicted observation_{t+1}
The model predicts the next observation conditioned on the current latent state and the executed action.
The model updates continuously while the environment runs.
Training loop:
(obs_t, action, obs_t+1)
↓
replay buffer
↓
mini-batch sampling
↓
gradient update
This enables live learning while interacting with the environment.
Prediction accuracy is evaluated using multiple metrics:
- MSE - Mean Squared Error
- MAE - Mean Absolute Error
- PSNR - Peak Signal-to-Noise Ratio
- SSIM - Structural Similarity
- Histogram Correlation
- RGB Channel Error
Metrics are tracked over time to measure learning progress.
The interface visualizes:
- prediction quality
- training loss
- gradient norms
- learning rate
- replay buffer size
- latent state statistics
- prediction error heatmaps
This makes the learning dynamics visible in real time.
The model can run in Pure Dream Mode, where it generates predictions based only on its own previous predictions rather than real observations.
This allows exploration of the internal generative dynamics of the learned world model.
Clone the repository:
git clone https://github.com/yourusername/mini-dreamer.git
cd mini-dreamer
Install dependencies:
pip install -r requirements.txt
Run the interactive environment:
python play_learn.py
WASD move
H toggle error heatmap
L toggle online learning
P pure dream mode
R reset latent state
F5 save model checkpoint
Q quit
python play_learn.py --ckpt model.pt --seed 7 --scale 6
Arguments:
--ckpt model checkpoint
--seed environment seed
--scale rendering scale
Mini Dreamer implements a simplified world model architecture inspired by Dreamer-style systems.
Core components:
Environment
↓
Observation Encoder
↓
Latent Representation (z)
↓
Recurrent State Space Model (RSSM)
↓
Decoder
↓
Predicted Observation
The recurrent latent state captures temporal dynamics of the environment, enabling the model to predict future observations conditioned on actions.
mini-dreamer
play_learn.py
interactive environment + online training
train.py
world model training utilities
world_gen.py
procedural environment generation
assets/
demo video and preview image
checkpoints/
saved model weights
The model optimizes a combined objective:
L = reconstruction_loss + β · KL_divergence
Where
- reconstruction loss = L1 prediction error
- KL divergence regularizes the latent representation
This encourages learning a compact predictive latent space.
Potential extensions include:
- reinforcement learning agents
- curiosity-driven exploration
- transformer-based world models
- larger environments
- dataset logging for offline training
- multi-agent simulations
If you use this project in research or experiments, please cite:
@software{mini_dreamer,
author = {Tomasz Water},
title = {Mini Dreamer: Interactive World Model with Online Learning},
year = {2026},
url = {https://github.com/yourusername/mini-dreamer}
}
MIT License
Tomasz Wietrzykowski