Update architecture docs with state persistence and uncertainty
Browse files- docs/architecture.md +2 -0
docs/architecture.md
CHANGED
|
@@ -37,6 +37,8 @@ Grid-JEPA combines JEPA self-supervised learning with world model dynamics for A
|
|
| 37 |
- GRU sequence model
|
| 38 |
- Discrete latent encoder with straight-through gradients
|
| 39 |
- Imagination rollouts for planning
|
|
|
|
|
|
|
| 40 |
|
| 41 |
## 4. Hierarchical Controller
|
| 42 |
|
|
|
|
| 37 |
- GRU sequence model
|
| 38 |
- Discrete latent encoder with straight-through gradients
|
| 39 |
- Imagination rollouts for planning
|
| 40 |
+
- **CRITICAL: State persists across levels within an environment** — Level 3 requires knowledge from Level 1-2
|
| 41 |
+
- Uncertainty-aware prediction for hypothesis revision (when predictions fail, update the model)
|
| 42 |
|
| 43 |
## 4. Hierarchical Controller
|
| 44 |
|