| --- |
| license: apache-2.0 |
| library_name: torch |
| language: en |
| pipeline_tag: reinforcement-learning |
| tags: |
| - world-model |
| - latent-dynamics |
| - defect-regularization |
| - continuous-representation |
| - finite-dynamical-system |
| --- |
| |
| # AQARION Defect World Model |
|
|
| A foundational world model that learns structured latent transitions via the **AQARION Defect Regularizer**, a differentiable penalty that minimizes topological defects in latent space. Designed for environments with continuous state spaces and structured dynamics. |
|
|
| ## Model Description |
|
|
| This model combines: |
| - A standard encoder-decoder architecture for reconstruction. |
| - The **AQARION Defect Loss**, which regularizes latent transitions by penalizing misalignment between encoded states and their predicted evolution. |
|
|
| The loss encourages the latent space to evolve smoothly, reducing "defects" that disrupt downstream planning or control. |
|
|
| ## Intended Use |
|
|
| - Learning compact, structured world models from high-dimensional observations. |
| - Environments with toroidal or cyclic dynamics (e.g., robotics, climate, physics simulation). |
| - Research into latent space topology and structured regularization. |
|
|
| ## Limitations & Ethical Considerations |
|
|
| - The defect loss assumes a fixed number of clusters; performance may degrade if dynamics are highly non-stationary. |
| - Not intended for safety-critical deployment without additional verification. |
| - Latent interpretability is limited—clusters are emergent, not semantically labeled. |
|
|
| ## Training |
|
|
| - **Data**: High-dimensional continuous observations with temporal structure. |
| - **Optimizer**: Adam (lr=1e-3) |
| - **Batch size**: 256 |
| - **Latent dim**: 16 |
| - **Clusters**: 6 |
| - **Defect weight (β)**: 10.0 |
| - **Seed**: 369 (for reproducibility) |
|
|
| ## Usage |
|
|
| ```python |
| |
| from aqarion_world_model import AQARIONWorldModel |
| |
| model = AQARIONWorldModel(obs_dim=128, latent_dim=16, num_clusters=6) |
| |
| recon, defect_loss, clusters = model(obs_t, obs_t_next) |
| |
| --- |
| |
| @misc{aqarion2026defect, |
| |
| author = {AQARION}, |
| |
| title = {AQARION Defect World Model}, |
| |
| year = {2026}, |
| |
| publisher = {Hugging Face}, |
| |
| url = {https://huggingface.co/aqarion/aqarion-defect-world-model} |
| |
| } |
| --- |
| license: apache-2.0 |
| library_name: torch |
| language: en |
| pipeline_tag: reinforcement-learning |
| tags: |
| - world-model |
| - latent-dynamics |
| - defect-regularization |
| - continuous-representation |
| - finite-dynamical-system |
| --- |
| |
| # AQARION Defect World Model |
| |
| A foundational world model that learns structured latent transitions via the **AQARION Defect Regularizer**, a differentiable penalty that minimizes topological defects in latent space. Designed for environments with continuous state spaces and structured dynamics. |
| |
| ## Model Description |
| |
| This model combines: |
| - A standard encoder-decoder architecture for reconstruction. |
| - The **AQARION Defect Loss**, which regularizes latent transitions by penalizing misalignment between encoded states and their predicted evolution. |
| |
| The loss encourages the latent space to evolve smoothly, reducing "defects" that disrupt downstream planning or control. |
| |
| ## Intended Use |
| |
| - Learning compact, structured world models from high-dimensional observations. |
| - Environments with toroidal or cyclic dynamics (e.g., robotics, climate, physics simulation). |
| - Research into latent space topology and structured regularization. |
| |
| ## Limitations & Ethical Considerations |
| |
| - The defect loss assumes a fixed number of clusters; performance may degrade if dynamics are highly non-stationary. |
| - Not intended for safety-critical deployment without additional verification. |
| - Latent interpretability is limited—clusters are emergent, not semantically labeled. |
| |
| ## Training |
| |
| - **Data**: High-dimensional continuous observations with temporal structure. |
| - **Optimizer**: Adam (lr=1e-3) |
| - **Batch size**: 256 |
| - **Latent dim**: 16 |
| - **Clusters**: 6 |
| - **Defect weight (β)**: 10.0 |
| - **Seed**: 369 (for reproducibility) |
| |
| ## Usage |
| |
| ```python |
|
|
| from aqarion_world_model import AQARIONWorldModel |
|
|
| model = AQARIONWorldModel(obs_dim=128, latent_dim=16, num_clusters=6) |
| |
| recon, defect_loss, clusters = model(obs_t, obs_t_next) |
