cuOpt Pipe Router

An NVIDIA Omniverse extension that routes pipes through 3D environments while avoiding obstacles. Uses NVIDIA Warp for GPU mesh voxelization and NVIDIA cuOpt for GPU-accelerated route optimization.

Quick start

1. Start the cuOpt server

docker run --gpus all -it --rm -p 5001:5000 nvidia/cuopt:26.4.0-cuda13.0-py3.13

2. Install the extension in Omniverse

1. Open USD Composer (or Isaac Sim)
2. Go to Window > Extensions > gear icon
3. Add this repo's path as a search path
4. Search for cuOpt Pipe Router and enable it

3. Create a scene and route

1. Click Simple Scene or Engine Bay to load a preset
2. Drag the colored spheres to position pipe start/end points
3. Adjust resolution, clearance, and bend penalty
4. Click Route All Pipes

Scenes

Simple scene

Five cube obstacles with three pipe routes. Good for understanding the system.

Bend penalty = 0 (shortest path)	Bend penalty = 20 (fewer bends)

Engine bay

Loads a 3D scanned engine bay from assets/engine_bay.usdc with three pipe routes (coolant, oil line, AC line). Works with any imported mesh, just drop it under /World/Obstacles.

How it works

Scene geometry (any mesh under /World/Obstacles)
        |
        |  Warp GPU kernel: signed distance query per grid cell
        |  Inside mesh or within clearance = blocked
        v
Occupancy grid (3D voxel grid, red = blocked, green = free)
        |
        |  Build CSR waypoint graph from free cells
        |  26-connected neighbors, edge weight = distance + bend penalty
        v
CSR graph --> JSON --> POST to cuOpt server
        |
        |  cuOpt solves optimal route on GPU
        |  Returns waypoint-level path
        v
Smooth with Catmull-Rom spline, generate tube mesh

Pipes are routed sequentially. Each completed pipe becomes an obstacle for the next one.

Debug visualization

Enable the checkboxes under Debug Visualization to see what the solver sees:

• Red cubes = blocked cells (obstacle geometry + safety clearance)
• Green dots = free cells near obstacles (the nodes cuOpt can route through)

Parameters

Parameter	What it does
Grid Resolution	Cells along the longest axis. Higher = more accurate voxelization, slower solve. 30 is fast, 60+ for detailed meshes.
Safety Clearance	Minimum distance (in scene units) between pipes and obstacles.
Tube Radius	Visual thickness of the generated pipe.
Bend Penalty	0 = shortest path with diagonal cuts. Higher values prefer axis-aligned segments with fewer bends.

What does cuOpt optimize?

Each pipe is a single-vehicle routing problem on the waypoint graph. cuOpt minimizes total edge weight:

weight = euclidean_distance + bend_penalty * max(0, axes_changed - 1)

The graph is sent as CSR (Compressed Sparse Row) format:

• offsets[i] to offsets[i+1] = edge range for node i
• edges[j] = destination node of edge j
• weights[j] = cost of edge j

Project structure

config/extension.toml
assets/                       scene assets (engine_bay.usdc)
docs/
omni/cuopt/
    extension.py              extension lifecycle, multi-pipe orchestration
    warp_voxelizer.py         GPU voxelization via Warp SDF queries
    cuopt_solver.py           CSR graph builder + cuOpt REST client
    pathfinding.py            occupancy grid, path smoothing
    scene_builder.py          USD geometry: obstacles, markers, tubes
    debug_viz.py              occupancy grid + free cell visualization
    ui/panel.py               omni.ui panel

Requirements

• NVIDIA GPU
• NVIDIA Omniverse (works with any kit application, I tested it on IsaacSim)
• omni.warp extension enabled
• cuOpt 26.4.0+ Docker container