Skip to content

Project Roadmap

This roadmap outlines the upcoming milestones for DSG‑JIT, spanning stability, features, optimization, and long‑term research directions.
It is divided into phases with clear goals, deliverables, and stretch objectives.

🚀 Phase 1 — Core Stabilization (Completed)

Status: ✔️
Summary: Completed foundational work on optimizer, SE(3) manifold, voxel grid operators, scene graph, residuals, and testing suite.

Deliverables: - JIT‑friendly factor graph engine
- Full SE(3) geodesic math and differentiable odometry
- Voxel smoothness, point observation, and multi‑parameter wrappers
- Unified Gauss‑Newton solver
- 26/26 passing tests

🔧 Phase 2 — Differentiable Scene Graph (Completed)

Status: ✔️
Introduced the world model, scene graph relations, entity system, and DSG‑based optimization hooks.

Deliverables: - Relational scene graph (parent/child, rigid attachments, etc.) - Voxel + SE(3) hybrid factors
- Differentiable world model component
- Param‑learnable factors for odom & voxel observations

🧪 Phase 3 — Experiments & Validation (Completed)

Status: ✔️
All algorithmic experiments defined, executed, and reproduced: - Learnable type weights
- Learnable odom measurements
- Multi‑voxel observation learning
- Hybrid SE3 + voxel joint learning (hero experiment)

📈 Phase 4 — Benchmarks & Performance (Completed)

Status: ✔️
Three official benchmarks implemented: - Pure SE3 factor graph
- Voxel grid smoothness chain
- Hybrid SE3 + voxel chain

JIT speedups: 31–7000× depending on graph size.

🤖 Phase 5 — Real‑World Sensors & SLAM Integration (Completed)

Integration of DSG‑JIT into full robotics pipelines.

Planned: - Real LIDAR factor
- RGB‑D depth factor
- Visual landmarks
- Camera intrinsics/extrinsics calibration via DSG
- Data loaders for KITTI / TUM RGB‑D

Stretch: - IMU pre‑integration
- Multi‑robot DSG fusion

📚 Phase 6 — Public Documentation (Completed)

Status: 🟡
DSG-JIT current development stage.

Remaining tasks: - Polish docs (architecture, API, examples, benchmarks)
- Generate gallery diagrams
- Ensure docs build cleanly under GitHub Pages
- Add narrative tutorial series

Stretch: - Animated diagrams showing optimization steps
- Interactive code sandboxes

🧩 Phase 7 — Packaging & Distribution (Completed)

Status: ⏳

Planned deliverables: - pip install dsg-jit
- Versioned releases + changelog
- Improved import layout
- Automated lint + format + test pipeline
- Pre‑commit hooks
- GitHub Actions for: - type-check
- tests
- benchmark snapshot
- docs deploy
- ROS2 wrapper package

Stretch: - Optional CUDA/XLA GPU acceleration
- Wheels for Mac, Linux, Windows

🧬 Phase 8 — Research Extensions (Long‑Term)

  • DSG‑based reinforcement learning
  • Learned Jacobian priors
  • Generative world‑model layers

Potential publications: - Differentiable Scene Graph Optimization via JIT Factor Graphs
- Hybrid SE3–Voxel Graphs for Dense Reconstruction
- End‑to‑End Learnable SLAM via Multi‑Residual Differentiation

🏁 Phase 9 — 1.0 Stable Release (Future)

The first fully stable release of DSG‑JIT.

Requirements: - Complete documentation
- Full packaging
- All critical benchmarks validated
- Public examples + tutorials
- Long‑term support policy
- Optimization safety & performance guarantees
- Extensive DSG API for "plug-and-play" Scene Graph Rendering

🏗️ Phase 7 — Advanced DSL & Autogeneration (Planned)

A domain‑specific "DSG Modeling Language" for declarative factor-graph design.

Features: - YAML/JSON graph definitions
- Auto‑generated optimization graphs
- Auto‑differentiated residual templates
- Scenegraph compiler → JIT graph

Stretch: - Visual graph editor
- Drag‑and‑drop factor construction UI
- "Graph debugger" visualization

Contributing

Contributions are welcome!
Upcoming needs: - More tests
- More factor types
- Benchmark expansions
- Doc improvements
- Tutorials + examples