← Voxel World Model

Roadmap

What is shipped, what comes next, and how each phase builds on the authoritative Adaptive Voxel World — without rewriting the architecture.

Invariant across every future milestone

The Adaptive Voxel World remains the single source of geometric truth. Perception proposes observations; the Evidence Engine decides belief; Components preserve identity; the SVO accelerates queries; the runtime coordinates incremental recompiles; every visualization is a derived view. Later milestones add systems — they do not replace the world.

Shipped through V30Kernel v1.0 declared. This is the last planned kernel milestone. Prior: V29 trust/ops, V28 sync, V27 query, V26 kernel wiring. V13 streaming / global frame post and V15 Part 2 remain open polish. Future work versions as independent platform tracks (Perception, Simulation, Neural, Developer, Robotics, FRAME, Visualization) — not V31/V32.

Direction after v1.0: VWM becomes a FRAME spatial computation dApp

A lead-architect review (migration proposal) reaches an uncomfortable conclusion: the V26–V30 kernel is the FRAME platform kernel. Transactions, replication, trust, snapshots, invariants, and distributed history are FRAME primitives — so ~45–55% of the codebase should be deleted or handed to FRAME. What remains is a pure, deterministic spatial core: apply(state, op) → state', project(state) → FrameUnits, query(state, q) → result, with rendering externalized into Voxel View dApps.

This is a proposal, not shipped work — no code has changed and Kernel v1.0 stands as built. It reframes the platform tracks above as FRAME concerns and treats the migration as a deletion-led refactor guarded by the V30 golden-master determinism suite. Full write-up: VWM → FRAME migration →

Shipped Foundation (through V12)

EraMilestonesWhat it established
Capture → world Depth scan · real capture · TSDF v1–v5 · v7 Polycam · v8 feature voxels · v9 error-driven Reality → HAV / TSDF / adaptive voxels; hollow free space; extracted surface; multi-source fusion; measured error budgets; appearance decoupled from geometry
Acceleration & OS SVO · V10 World Engine · V11 Runtime Derived octree; universal upload; Components; revisions; evidence; dirty regions; scheduler; event bus; CompiledRepresentation
Perception → Gaussians V12 ReconstructionProvider stack; LingBot audited; pairwise ICP; chunked Gaussian scene + elliptical splat renderer (73k splats live)
Global mapping V14 SE(3) pose graph + Gauss-Newton; loop closure; landmarks/anchors/uncertainty; multi-capture fusion; incremental corrections via V11 runtime (8.99 m → 0 m loop)
Compiler kernel V15 Part 1 Single World Compiler Runtime; Artifact + Compiler contract; 5 plugins (gaussian/svo/mesh/semantic/lighting); transactional incremental compile; extensibility via register()
Visualization V16 Universal Visualization Engine; 17 artifact-only renderers; hot-switch never recompiles; multi-viewport; stack complete (world → compiler → viz)
Simulation V17 World alive over time; behaviors/triggers; tick pipeline through Runtime; compiler auto_wire to ChunkInvalidated (door → 768→764 splats); deterministic replay
Planned Foundation: V12 Gaussians

V13 — Production Gaussian Engine

Turn the V12/V15/V16 Gaussian path into a production observation pipeline: streaming rasterization plus optional local and global frame enhancers — without ever letting AI become the world authority.

Adaptive Voxel World │ ▼ ══════════════════════════ World Compiler ══════════════════════════ │ ▼ Gaussian Scene Artifact │ ▼ ══════════════════════════ Gaussian Rasterizer ══════════════════════════ │ ▼ Gaussian Frame + Metadata │ ┌────┴────┐ ▼ │ Local AI │ (materials, textures, fine detail) │ │ ▼ │ Enhanced │ └────┬────┘ ▼ Final Display Frame │ ▼ Global AI (optional) (HDR, grading, atmosphere, bloom, DOF, temporal, style) │ ▼ Monitor

Goals

Definition of done

Depends on

V12 Gaussian scene · V15 compiler artifacts · V16 visualization engine · homepage / paper enhancement path. See also Frame Enhancement on the homepage.

Shipped — 88/88 tests

V14 — Complete Registration & Mapping ✓

Delivered. Maintained SE(3) pose graph with Gauss-Newton optimization, loop closure, landmarks, anchors, uncertainty, multi-capture fusion, and incremental voxel corrections through the V11 runtime. Measured: 12-node loop drift 8.99 m → 0.0 m in 70 ms. Full report: updates.html#v14.

Remaining follow-ups (honest): analytic Jacobians + sparse Cholesky for large graphs; higher confidence gate for tiny-overlap ICP; in-viewport 3D pose-graph / covariance modes and interactive constraint editor; live GPS/IMU/ROS/LiDAR adapters when hardware is present.

Shipped Part 1 — 105/105 tests

V15 — World Compiler Framework ✓ Part 1

Part 1 delivered. Single World Compiler Runtime owns all derived-artifact compilation; five built-in plugins wrap real Gaussian / SVO / mesh / semantic / lighting logic; incremental, transactional, cached, deterministic-where-applicable, and extensible without touching the runtime class. Full report: updates.html#v15.

Part 2 remaining: route the viewer’s live Gaussian/SVO rendering through world.compiler artifacts (today dual-path with V11 runtime.compilers); parallel multi-worker execution; richer queue / dependency-graph / timeline panels; stock compilers for navmesh / physics hulls / neural packs as they land.

Shipped — 118/118 tests · stack complete

V16 — Universal Visualization Engine ✓

Delivered. One Visualization Engine; 17 renderers consume V15 compiler artifacts only; hot-switch never recompiles; multi-viewport; cameras / layers / selection / screenshots / recording through one API. Caps the major subsystem stack. Full report: updates.html#v16.

Remaining polish: route Three.js live pixel drawing fully through /api/viz/render for every mode; ship stylized renderers (photoreal / LEGO / anime / thermal) as plugins over the same artifacts.

Shipped — 135/135 tests · world alive

V17 — World Simulation ✓

Delivered. Simulation evolves the authoritative world through time with no new foundational layer. Behaviors / triggers / tick / time control; changes flow Runtime → Evidence → World → Compiler (auto) → Visualization. Proven: door open auto-recompiles gaussian 768→764. Full report: updates.html#v17.

Follow-ups: richer physics on the same Behavior interface; browser play/pause/step strip; bump runtime revision on all world edits (same as sim).

Shipped · 150/150 tests

V18 — Gaussian Enhancement Pipeline

Delivered. Enhances the Gaussian render itself (not full-frame display post). Flow: compiler Gaussian artifact → CPU splat rasterizer → GaussianImagePacket → visible-region metadata → 32-D embedding → chainable enhancers → enhanced image. Voxel world untouched; disabled = deterministic passthrough. Full report: updates.html#v18.

Honest limits: classical-CV path is what runs without GPU weights; AI enhancers are guarded stubs; CPU reference rasterizer ≠ browser WebGL; display-mode wiring for the enhanced frame is a follow-up.

Note: Digital Twin / Spatial Database proceeds under platform tracks (Perception · Simulation · FRAME), not further kernel milestones.

Shipped · 164/164 tests

V19 — Gaussian Evolution Engine

Delivered. Continuously improves Gaussian scene data (not pixels) asynchronously: compiler scene → optimizer passes → cached revisions → best version served transparently via best_arrays(). Complements V18. Hallway quality 0.65→1.0. Voxel world untouched; render never blocked. Full report: updates.html#v19.

Honest limits: classical splat processing is what runs without GPU weights; quality is a splat-data heuristic; predictive camera-motion prefetch is scaffolded only.

Shipped · 216/216 tests

V20 — Reality Capture Pipeline

Delivered. Unified doorway for all reality intake: sources emit ObservationPacket → Registration + Evidence → voxels. Never writes voxels directly. Synthetic live run; record→identical-replay; guarded hardware stubs. Full report: updates.html#v20.

Shipped · 231/231 tests

V21 — Reactive World Pipeline

Delivered. Runtime event bus + reactors: one world edit cascades VoxelModified → … → GaussianFrameEnhanced with zero orchestration code. Deterministic record/replay; backpressure 0. Full report: updates.html#v21.

Honest limits: V11 mark_dirty still schedules compile internally; reactors stay lightweight; LoopbackChannel is in-process.

Shipped · 246/246 tests · real CUDA

V22 — Neural Enhancement Backends

Delivered. Interchangeable neural optimization backends (never the renderer). Real-ESRGAN RRDBNet x4 runs on the RTX 2060 SUPER (device=cuda); evolution drives them via registry only. VRAM tiling + CPU fallback; FLUX/SwinIR/HiDream guarded until weights land. Full report: updates.html#v22.

Honest limits: scene update is neural-gated colour-detail transfer (detail_gain), not full splat inverse-rendering.

Shipped · 260/260 tests

V23 — Semantic World Engine

Delivered. Meaning decoupled from geometry: entities own hypotheses, attributes, and relationships; only reference authoritative voxels. One evidence path; competing hypotheses with confidence; directional relationships with auto-inverse; independent persistence. Full report: updates.html#v23.

Honest limits: HeuristicProposer is shape-based not deep recognition; search() is lexical; no bundled perception model — real detectors attach via propose(source="ai").

Shipped · 276/276 tests

V24 — Persistent World Identity

Delivered. Objects keep persistent UUIDs across observations: multi-signal probabilistic matcher, accumulated confidence, immutable history/versions/lifecycle, change detection, and state_at/replay/compare without mutating the present. Full report: updates.html#v24.

Honest limits: matcher is a transparent weighted heuristic (not learned embeddings); auto merge/split is explicit API; capture→identity auto-wiring is next integration.

Shipped · 289/289 tests · stack complete

V25 — World Reasoning Engine

Delivered. Fourth layer of the stack: evidence-backed inference, classical rules, causal graphs, competing hypotheses, explanations, and hypothetical predictions — never mutates the authoritative world. Caps reality → meaning → continuity → inference. Full report: updates.html#v25.

Honest limits: classical rules only (no bundled AI); disappearance priors labelled as such; prediction is persistence projection; AI reasoner is a plugin interface.

Shipped · 311/311 tests

V26 — Integrated World Kernel

Delivered. V20–V25 compose automatically: atomic transactions, event-driven kernel reactors, whole-world snapshot save/restore with identical authoritative checksum, 7 executable invariants. One observation → whole stack; no manual downstream calls. Full report: updates.html#v26.

Honest limits: no incremental snapshots; learned perception not added; product e2e now fixed (was shadowed reconstruction import). Cross-layer query delivered in V27.

Shipped · 329/329 tests

V27 — World Services & Query Engine

Delivered. The world is a service: world.query is the single app interface — spatial/temporal/semantic/identity/reasoning/graph families plus cross-layer explain (one WorldExplanation). Full report: updates.html#v27.

Honest limits: ask() is a keyword parser not NLP; planner is index-vs-scan; large-world path planning out of scope.

Shipped · 341/341 tests

V28 — Distributed World Replication & Synchronization

Delivered. Many replicas, one logical world — converge through validated transactions, vector clocks, explicit conflicts, checkpoints, offline workflow, and abstract transport (mem + file). Full report: updates.html#v28.

Honest limits: HTTP/WebSocket not built (Transport contract only); state merge by uuid; discovery abstract; hallway snapshot is not in the replication log (live demo uses fresh transactional replicas).

Shipped · 355/355 tests

V29 — Distributed Kernel Hardening & Trust

Delivered. Replicate operations, not conclusions — local recompute, HMAC signatures, trust roles, provenance, versioned policies, compaction / partial sync / anti-entropy. Full report: updates.html#v29.

Honest limits: HMAC default (Ed25519 arrived in V30); policy migration was detect-only until V30 registry; anti-entropy was on-demand until V30 scheduled tick.

Shipped · Kernel v1.0 · 366/366 tests

V30 — Kernel Finalization & Long-Term Stability

Delivered — last kernel milestone. Observation-driven identity, Ed25519 public-key + federation, policy evolution, scheduled anti-entropy, health score, canonical regression (5 worlds ALL OK), governance + KERNEL_API + KERNEL_V1.0 declaration. Full report: updates.html#v30.

Honest limits: pure-Python Ed25519 is slow (HMAC remains fast default); anti-entropy is a throttled tick; concrete policy migrations are app-registered.

Horizon · independent tracks

Platform tracks — version above the kernel

No more V31, V32… kernel milestones. The kernel stays at v1.0 while these tracks version independently (see docs/KERNEL_V1.0.md):

Kernel v1.0 (stable substrate) Perception 0.x — real sensing → authoritative observations Simulation 0.x — living world (physics / lighting / HVAC / traffic / …) Neural World 0.x — Gaussian / radiance / appearance learning Developer 0.x — SDKs (Python / Rust / C++ / TS) · REST/gRPC · plugins Robotics 0.x — localization / navigation / fleet sync via the world FRAME 0.x — spatial OS (apps / permissions / workspaces / automation) Visualization 0.x — editor / timeline / replay / trust+replication inspectors

Themes still on the horizon

Non-goals (still)

How to read this roadmap

V1–V9 Capture & adaptive voxels (measured) V10–V11 World OS + runtime + evidence V12 Perception providers + Gaussians (scaffolded) V14 ✓ Global registration & mapping (shipped) V15 ✓P1 Compiler kernel (shipped) V16 ✓ Visualization engine (shipped · stack complete) V17 ✓ World simulation (shipped · world alive) V18 ✓ Gaussian enhancement pipeline (classical-CV live) V19 ✓ Gaussian evolution engine (async splat cache) V20 ✓ Reality capture pipeline (unified ObservationPacket) V21 ✓ Reactive world pipeline (event bus + reactors) V22 ✓ Neural enhancement backends (RealESRGAN on CUDA) V23 ✓ Semantic world engine (meaning decoupled from geometry) V24 ✓ Persistent world identity (timeline + replay) V25 ✓ World reasoning engine (inference · causal · explain) V26 ✓ Integrated world kernel (tx · events · snapshot · audit) V27 ✓ World services & query (cross-layer explain · graph) V28 ✓ Distributed replication (vector clocks · conflicts · checkpoints) V29 ✓ Distributed trust (operations · recompute · HMAC · provenance) V30 ✓ Kernel v1.0 (regression · Ed25519 · governance · FINAL kernel) │ ▼ V13 Production Gaussian streaming + global frame post V15 P2 Viewer migration through /api/viz Tracks Perception / Simulation / Neural / Developer / Robotics / FRAME / Viz

Shipped milestones earn their place with tests and numbers. Planned milestones above are the intended product arc; dates are intentionally omitted until a milestone opens with a definition-of-done and an honest environment report — the same standard used for V8–V12.

Shipped updates → Technical paper → Benchmarks →