NULLVECTOR

NULLVECTOR is a 2.5D cellular world simulation built around physical creatures, ecology, evolution, and neural rendering. The short version is:

Powder Game + Spore + Rain World + Caves of Qud.

The current build is a playable research prototype. Creatures are assembled from cells, organs, skeletons, muscles, fluids, and articulated appendages. They move through a top-down world while remaining vertically oriented. They can feed, fight, heal, reproduce, mutate, grasp materials, lose limbs, and die.

The long-term target is one recurrent action-conditioned DiT with a continuous VAE decoder: controls and world state in, playable frames and causal state out. We are getting there in stages instead of pretending one model already does everything.

Current build

• Five organism families: humanoid, animalian, plantlike, anomaly, and machine.
• Cellular bodies with organs, circulation, fluids, damage, healing, scarring, severing, and death.
• Grounded musculoskeletal locomotion with planted feet, limb constraints, and family-specific movement.
• Articulated grasping, feeder contact, throwing, recoil, elevation, shadows, bounce, roll, and thud material responses.
• Family-specific acquisition: hands, mouths, roots, phase fields, and machine tools.
• Metabolism, predation, reproduction, mutation, resources, construction, settlements, and persistent streamed regions.
• Validated semantic maps across six visual themes.
• Native Godot 4 runtime with Python used for training, evaluation, and asset generation.

The simulation is broad but still early. Creature construction and physical interaction are the strongest parts. Ecology, societies, cities, and planetary scale exist as scaffolds and research systems rather than a finished game.

Selected results

Playable neural foundation

The live nature simulation now batches current posed and damaged cells through the continuous organism VAE while neural controllers drive locomotion, physiology, feeding, behavior, colonies, societies, and forecasts.

Grounded motion and feeding	Five feeding strategies

Ballistic throwing	True limb severing

Neural cellular locomotion	Anatomical VAE motion

Continuous neural cell raster

Recurrent world rollout: rollout decoder vs foreground-aware decoder

More results are in the output gallery.

Neural world synthesis	Raw topology and safety repair

Neural city layout

The current world-synthesis path composes a decoder-coupled neural topology prior with the accepted neural decal/prop selector. It produces all six biome regions in 6.9 seconds with about 110 MiB peak reserved VRAM. Raw maps connect required mission points in 83.3% of the six-theme sentinel set and support the full agent radius in 50%; the safety compiler changes 5.45% of cells on average. That compiler is still authoritative, so this is an experimental neural world foundation rather than a claim that topology is fully learned.

What is neural

NULLVECTOR currently uses an ensemble. Deterministic systems remain wherever they are still the safer or better authority.

System	Current authority
Creature fields and identity	Neural categorical generation with validated anatomy fields
Sprite rasterization	Promoted continuous cell VAE; current posed cells in, 96x96 RGBA out
Grounded motion	Neural muscle/contact feedback inside physical constraints
Limb pose and grasping	Neural inverse-muscle and grasp controllers inside an articulated solver
Local cell dynamics	Promoted causal cellular NCA with organ-ablation and rollout gates
Maps	Procedural topology authority with neural topology and decoration models under evaluation
Ecology and societies	Neural behavior, colony, society, timeline, counterfactual, and spatial city-layout specialists inside the causal scaffold
World frames	Recurrent Action-DiT + adapted continuous VAE change compositor; live in the Python nature stage, not yet the native Godot authority

The mobile local-cell rule now distills the 10M-parameter causal physiology teacher into 492k parameters and a 2.0 MB ONNX graph. Across an untouched five-family holdout it reaches 0.981 response cosine over 32 recurrent steps, matches circulation, respiration, digestion, and neural-ablation direction, and keeps internal state inside the organism while surface fluid diffuses into the world. Report and model artifacts are checked in with the rollout comparison below.

Android preview

The Android preview runs the recurrent action core, cellular physiology NCA, grounded controller, grasper policy, ecology, and high-level world ensemble in one 2.5D habitat. It now has a proper setup screen, independent movement and aim sticks, physical grasp/feed/throw interactions, and trait-gated machine and anomaly projectiles.

Its visible world is still the deterministic cell scaffold. The VAE currently feeds cell appearance and diagnostic frames; it does not yet decode the whole viewport. The target renderer is controls + recurrent world state -> visual latent field -> VAE viewport, with ordinary graphics limited to menus, HUD, accessibility, and debug overlays.

Playable habitat	Optional neural debug

The debug view identifies the old blurry image correctly as a world-latent decoder probe. It is hidden during play and only runs at debug cadence.

Composite Build 2 loads 67.7M parameters across the Action-DiT, world VAE, exact-parent pixel refiner, actor-state student, organism cell VAE, and causal physiology model. The larger 13-specialist teacher ensemble is hash-closed for reverse distillation.

The first recurrent action-frame student now beats exact frame persistence by 10.43% on its untouched cellular world. Its decoder adaptation cuts cellular reconstruction MAE by 91.0% while keeping the original encoder and latent contract frozen. This result is promoted as a tested component, not yet as the authoritative world simulation. It is callable in the nature stage as the live action-conditioned future view.

The rollout-aware decoder is trained on the recurrent model's predicted latent distribution, not only exact VAE encodings. On an untouched world it reduces rollout-frame MAE from 0.05742 to 0.01677 (70.8%) while also improving exact frame reconstruction. It removes most duplicated-body trails, although long rollouts remain softer than the target and are not yet game-quality.

The foreground-aware successor improves rollout MAE another 5.4%, foreground MAE 7.24%, and edge error 1.72%. The packaged 57.0M-parameter recurrent + VAE pipeline runs at 76.7 decoded 256x256 frames/s on the development GPU, with 432 MiB peak reserved VRAM and 228 MB of checkpoint artifacts.

The live nature stage targets 30 rendered frames/s. Grounded organism and appendage physics now advance at 30 Hz while the causal world and physiology advance at 15 Hz. Causal physiology advances four organisms per world tick through a batched neural update; slower ecology and society decisions stay on their own cadence. The organism VAE decodes identity-stable cell appearance once, then physics moves those neural-authored cells without rerasterizing an entire body per pose. Peak reserved VRAM is about 770 MiB.

The width-96 neural topology prior now trains through the frozen VQ decoder, not only against token IDs. A 200-step semantic fine-tune adds rare-token balancing, required-point walkability, decoded condition matching, categorical sharpness, and scale-aware differentiable agent-width reachability. The selected 500-step semantic checkpoint reaches 70.8% raw required-route connectivity and 41.7% agent-radius connectivity across all 24 held-out size/theme maps. Its six-biome 32px sentinel slice reaches 83.3% and 50%, with 5.5% mean deterministic repair. Hazard placement remains a separate neural-decorator responsibility. This is a measurable improvement and a useful failure boundary—not a promoted world generator.

The next recurrent student carries both visual and 128-feature organism state through contiguous rollouts. On the untouched world it beats its unchanged parents by 2.36–2.79% at every 4–32 step horizon, while beating initial-frame persistence by 9.10–12.33%. Its validated stream runtime is now the live F6 diagnostic path at 9.4–10.6 recurrent steps/s and 490 MiB peak reserved VRAM. The historical training set included diagnostic overlays, so this model is kept out of clean student view until it is retrained on the new overlay-free targets.

The overlay-free successor corpus now contains 2,376 contiguous frames across six worlds and all 22 action classes. Its first clean recurrent student beats the historical model at every 1–32 step horizon and beats frame persistence at 8, 16, and 32 steps. It still misses the four-step persistence gate, so it remains an experimental checkpoint rather than runtime authority.

The physical projector still enforces bone lengths, attached roots, planted contacts, collision safety, and feeder contact. That is intentional. The callable ensemble is now integrated into the nature stage; the next phase is broader rollout evaluation and reverse distillation into a recurrent student.

Run the game

Requirements:

• Godot 4.3
• Python 3.12 for forge and training tools
• A CUDA-capable GPU for production training; the native game does not require Python or CUDA

Open game/project.godot in Godot, or launch it directly:

C:\path\to\Godot_v4.3-stable_win64.exe --path C:\path\to\nullvector\game

The main scene is CreatureStage.tscn.

Controls:

• WASD — move
• Mouse — aim
• Left click — attack
• E — interact or assimilate
• Q — family utility
• F — build
• R — mutate
• Space — sprint
• Z/X/C/V/B — expressions and actions

The Python nature-stage demo uses F1 for a click-driven overlay panel. It only controls presentation and information: vision and sensed-target markers, anatomy and organs, integrity bars, labels, ecology links, settlements, sites, atlas, shadows, selection, evolution offers, mechanism telemetry, and the status HUD. F2 switches to the controlled organism's clean perception view: perfect awareness of its own cells, a broad hearing bubble, and its live aimed cone or radial sense. Seen map materials persist as last-known memory; unseen organisms and moving objects remain simulated but unrendered. L toggles the cone visualization; Shift+L toggles sensed-target markers. Neither key disables sensing. Teacher trajectories record the visibility field before diagnostic overlays are composited. These controls never pause or disable simulation, physics, AI, metabolism, damage, or ecology.

Thrown matter is physical and family-relative. Inert mineral or debris carries kinetic damage; matter the target can use instead repairs a bounded contact patch and transfers a small amount of energy. A single clump cannot repeatedly damage the same organism during one throw.

Forge and tests

Install the Python package dependencies, then run:

python -m pytest

Useful entry points:

# Generate semantic maps
python -m forge.maps generate --output outputs/maps_v2

# Render map art
python -m forge.map_art showcase `
  --map-sources outputs/maps_v2_forge_lab `
  --output outputs/map_art

# Replay the neural motion bank
python -m forge.multifield_style_neural_motion replay `
  outputs/multifield_style_neural_motion/motion_style_neural_manifest.json `
  --report outputs/multifield_style_neural_motion/verification_report.json

Most generated corpora, checkpoints, and evaluation banks are intentionally excluded from Git. Curated visual results live under examples/showcase/.

Architecture

game/                 Godot runtime and playable scenes
forge/                generators, models, trainers, evaluators, and compilers
shared/schema/        artifact and replay contracts
docs/                 subsystem design and validation notes
examples/showcase/    compact visual results for GitHub
outputs/              local generated artifacts and checkpoints

Every promoted neural component is evaluated against the causal scaffold. A model does not become runtime authority because its loss decreased; it must preserve anatomy, contacts, motion, damage semantics, and replay integrity.

Roadmap

1. Finish the neural creature foundation: morphology, rasterization, grounded locomotion, grasping, feeding, damage, and physiology.
2. Build a sustainable nature simulation with stable ecosystems, breeding, colonies, and mutation.
3. Expand into traits, equipment, construction, cities, societies, history, quests, and Caves of Qud-scale systemic variety.
4. Run the world as persistent patches on a zoomable planet with slower neural weather, migration, biome, and settlement updates.
5. Reverse-distill the proven ensemble into a recurrent action-DiT + VAE world model.
6. Optimize the finished desktop system for high-end mobile hardware.

The goal is not an AI-generated content demo. It is a real game and a proof that neural game systems can be deep, coherent, inspectable, and worth playing.

Documentation

• Creature-stage design
• Neural motion training
• Neural grounded target field
• Sprite field training
• Sprite evaluation
• Neural rig bridge
• Neural map systems
• Native workshop