Constraint-aware 3D routing of cables and pipes through automotive USD scenes, inside NVIDIA Omniverse. An Omniverse Kit extension (window name: PipeRouter) voxelizes the stage with NVIDIA Warp and hands the grids to a containerized solver, which finds the routes with NVIDIA cuGraph shortest paths on the GPU. Electric wires, CAN buses, AC lines and cooling circuits are all routed the same way, with different constraints.
| Piece | Where | What |
|---|---|---|
services/solver/piperouter_solver |
pure Python, GPU optional | grids → direction-aware weighted lattice (default: adaptive octree_lattice) → SSSP → fibre-neutre smoothing; all constraint math |
services/solver/piperouter_service |
Docker, cuGraph/GPU | FastAPI /solve and /solve_all; grids handed over via /dev/shm/piperouter |
exts/omni.piperouter |
Omniverse Kit | omni.ui panel, Warp voxelization, thermal/EM fields, USD tube authoring |
Built on NVIDIA Warp for GPU voxelization in-process in Kit, NVIDIA cuGraph for single-source shortest paths in the solver container, numpy and scipy for the solver core, FastAPI for the service, and OpenUSD with omni.ui for the extension. There is a scipy Dijkstra fallback throughout, so everything also runs without a GPU.
Hard constraints are collision and safety clearance (a global default or per-object
clearance tags), a thermal melt cutoff, waypoints, and pinned start/end headings from
either axis presets or a rotatable arrow gizmo. Soft constraints enter as weighted edge
cost: surface-hug, thermal, EM scaled by wire sensitivity, a bend penalty, and smoothing
strength. Wire types (cost, mass, diameter, bend radius, temperature rating) live in
wire_types.json.
# 1. Start the GPU solver (workstation with an NVIDIA GPU + Container Toolkit)
docker compose up --build -d
curl http://localhost:8000/health # {"status":"ok","backend":"gpu"}
# 2. Enable the extension in USD Composer / Isaac Sim:
# Window > Extensions > add this repo's exts/ to the search paths > enable omni.piperouterexts/omni.piperouter/docs/README.md has the full workflow: Route All, refining a single
wire with waypoints, locking it, then exporting a BOM. Service wiring is in
docker-compose.yml.
python3 -m venv .venv && .venv/bin/pip install -e services/solver
.venv/bin/pip install numpy scipy pytest fastapi "uvicorn[standard]" httpx usd-core warp-lang
# solver core + service
cd services/solver && ../../.venv/bin/pytest -p no:pqm -q # 130 tests
# extension headless logic (Warp voxelize, USD authoring, real-HTTP solve)
cd exts/omni.piperouter && ../../.venv/bin/pytest -p no:pqm -q # 88 testsThe GPU paths (cuGraph in the container, Warp in the extension) run on real hardware when it is present. The solver core falls back to scipy Dijkstra, so the test suites pass on a machine without a GPU.
Working today: the solver core, the GPU service, and a Kit extension covering Route All, per-wire refinement with waypoints (double-click a tube to drop one), bundles with shared trunks, the start/end heading gizmo, thermal/EM/clearance tagging including instanced CAD, buried-endpoint rescue, occupancy/thermal/EM overlays with per-wire debug views, session save/load into a single USD, and BOM export.
The default planner is the adaptive octree_lattice, which confines the search to a
corridor and is roughly an order of magnitude faster than the dense lattice at high
resolution. docs/benchmark_gpu_vs_cpu.md has the measured GPU-versus-CPU numbers per
stage and per resolution.
Possible next steps: design-space keep-in zones, drag-to-edit local re-solve, raceway corridors, a bundle-diameter formula, and pressure-drop checks for cooling circuits.