Skip to main content

Roadmap & Changelog

Changelog​

Phase 2 — model sharing, datasets, player preferences (2026-08-13)​

  1. Community model sharing — portable nv2-model-bundle export/import (/ai_export, /ai_import), NaN-sanitised and persisted on import.
  2. Training-dataset import — /ai_dataset <path> [epochs] trains the vegetation head on validated JSON datasets.
  3. Player-preference learning — per-class counters in the checkpoint (backward compatible), blended into training targets (30%).
  4. NaN checkpoint hardening — gradient clipping, sanitise-on-save, tolerant load; fixed a real corruption bug.

Suite: 96 passed / 0 failed (97 total incl. ignored release benchmark).

AI Phase 1 — vegetation intelligence (v1.0.0, complete)​

1. World height ceiling removed

  • CHUNK_H raised from 256 → 512 blocks
  • Taller mountains, deeper caves, prepared for unlimited-height streaming

2. 22 new vegetation block types (world/block.rs)

  • Flowers: rose, dandelion, tulips (4 colors), cornflower, allium, azalea
  • Water plants: lily pad, fern, fern plant, seagrass, tall seagrass, kelp
  • Decorations: small sticks, pebbles (3 variants), mossy cobble, vine, moss carpet

3. AI generator module (world/ai_generator.rs, NEW, ~600 lines)

  • TerrainAI MLP: 8→16→4, 320 parameters, ~1.2 KB
  • AISystem with background training thread
  • AIMessage channel for training progress / texture generation / vegetation decisions

4. AI integrated into world (world/mod.rs)

  • World::new_with_settings() spawns the AI system
  • ai_system + ai_receiver stored on World

5. Vegetation generation extended (world/vegetation.rs)

  • New place_ai_vegetation(): 3×3 cells, 8 terrain features, confidence threshold 0.5, biome-aware probabilities

6. Dependencies added (Core/Cargo.toml)

  • ndarray 0.15, rand 0.8, reqwest 0.11 (json), tokio 1 (full)

Behavior change (before → after)​

AspectBeforeAfter
FlowersRandom noise-basedAI-predicted placement
FernsNoneWet, shaded areas
Sticks/pebblesNoneNatural distribution
Learning—Continuous background training
FPS impact—None (<1%)

Roadmap — Phase 2​

Feature 1: Internet-based dataset integration — ✅ local part done​

  • ✅ Load + validate + train on JSON dataset files (/ai_dataset)
  • ⬜ Cloud dataset repository / API endpoints
  • ⬜ Live weather merge already runs (Open-Meteo, offline fallback)

Feature 2: Real-time AI texture generation​

  • ⬜ Generate textures procedurally with the network
  • ⬜ Integrate with the renderer's dynamic atlas
  • ⬜ Emit generated textures via AIMessage::TextureGenerated

Feature 3: Online learning from player actions — ✅ done (2026-08-13)​

  • ✅ Player choices tracked as preference counters in the checkpoint
  • ✅ Training targets blended with learned preferences (30%)
  • ✅ Privacy-safe: only local terrain features, no coordinates, no personal data

Feature 4: Cloud model sharing (multiplayer) — ✅ local part done​

  • ✅ Portable nv2-model-bundle export/import (/ai_export, /ai_import)
  • ⬜ Cloud upload/download endpoints / model marketplace
  • ✅ Versioned, self-describing checkpoint format

Feature 5: GPU acceleration​

  • ⬜ Parallelize training samples across GPU (batched forward/backward)
  • ⬜ Larger models with no frame-time cost

Feature 6: World & gameplay depth​

  • ⬜ Seasonal vegetation changes
  • ⬜ Multi-biome coordination
  • ⬜ Unlimited-height chunk streaming
  • ⬜ Dedicated gameplay audio system (movement-medium signals already tracked)
  • ⬜ Networking / multiplayer
  • ⬜ In-engine content editor

Known limitations (Phase 1)​

  1. No internet connectivity yet (Phase 2)
  2. Synthetic training data only
  3. Single-scale learning
  4. No model persistence across sessions
  5. Basic texture generation
Last updated on