VRP — GPU-Accelerated Field Service Optimizer

Real-time Vehicle Routing Problem solver for field service operations. Dispatches technicians to work orders using road-accurate travel times, skill matching, and workload balancing — powered by NVIDIA cuOpt on GPU.

---

What it does

• Assigns technicians to work orders optimally, minimising total travel time
• Respects time windows, shift hours, lunch breaks, and daily order limits
• Optional skill matching — only qualified technicians get assigned
• Optional workload balancing — caps total service time per technician so work is spread evenly
• Real road travel times via OSRM (not straight-line estimates)
• Interactive map with colour-coded routes and priority legends
• Demo data generator — enter a city name, choose order/technician counts, and the backend auto-generates realistic scenarios using real OSM locations (via Overpass → Nominatim → random fallback chain)
• Named scenario persistence — save, load, and delete named scenarios; stored as JSON files on the backend and survive container restarts

Architecture

Browser (Svelte + Leaflet)
    │  HTTP (same-origin in Docker, localhost:8000 in dev)
    ▼
nginx                         ← serves UI, proxies /vrp/* to API
    │
    ├── FastAPI (Python)       ← REST API, problem validation, result formatting
    │       │
    │       ├── cuOpt          ← NVIDIA GPU solver (DataModel → Solve)
    │       └── OSRM client   ← travel time matrix via HTTP
    │
    └── OSRM server           ← road routing engine (Docker container)

Quick start (Docker)

Prerequisites: Docker Desktop with NVIDIA GPU runtime, CUDA-capable GPU.

git clone https://github.com/grepjava/VRP.git
cd VRP
./setup.sh

Open http://localhost in your browser.

To use a different map region:

./setup.sh --map-url https://download.geofabrik.de/europe/germany-latest.osm.pbf

See SETUP.md for full installation and configuration options.

Development (without Docker)

See SETUP.md — Development workflow.

Solver settings

The UI exposes four cuOpt settings via the ⚙ Settings panel:

Setting	cuOpt API	Effect
Enforce skill matching	add_order_vehicle_match	Hard constraint — only assigns qualified technicians
Minimize fleet size	set_vehicle_fixed_costs	Penalises deploying extra vehicles
Balance workload	add_capacity_dimension	Caps total service time per technician — prevents overloading without pushing routes to late in the day
Custom time limit	SolverSettings.set_time_limit	More time = better solution quality
Drop return to base	set_drop_return_trips	Omit the final leg back to the depot (global toggle)

API

POST /vrp/optimize                  Run the solver
GET  /health                        Service health check
GET  /status                        Solver and GPU status
GET  /docs                          Interactive API docs (Swagger)

POST /vrp/generate-demo             Generate a realistic scenario for a given city
GET  /vrp/scenarios                 List saved scenarios (name, slug, counts, created_at)
POST /vrp/scenarios                 Save a named scenario (technicians + work orders as JSON)
GET  /vrp/scenarios/{slug}          Load a saved scenario
DELETE /vrp/scenarios/{slug}        Delete a saved scenario

Example request body: see SETUP.md — API usage.

Tech stack

Layer	Technology
Solver	NVIDIA cuOpt (GPU-accelerated VRP)
Routing	OSRM (Open Source Routing Machine)
API	FastAPI + Uvicorn
GPU memory	RAPIDS RMM
Data	cuDF (GPU DataFrames)
Frontend	Svelte 4 + Leaflet
Container	Docker Compose + nginx