This repository contains a benchmark dataset for variants of the Vehicle Routing Problem (VRP) generated using real-world road network data. It also includes scripts used to generate the dataset, enabling reproducibility and further extensions for new variants and ML applications.
The benchmark is designed for real-world (rich) VRP scenarios, incorporating multiple constraints and realistic city topologies. All instances are based on real geographic regions, while travel times are derived from road networks using routing services and OpenStreetMap data. The main goal of this benchmark is to enable fair comparison of algorithms on standardized yet realistic scenarios. The dataset covers a wide range of city structures and constraint combinations, allowing researchers to select instances that best match their specific application.
Currently, the benchmark focuses on DC-to-store distribution scenarios, characterized by a single depot and moderate time window constraints. This setup is motivated by practical insights obtained through collaboration with industry partners. Contributors are encouraged to extend the benchmark to other applications, such as urban delivery systems, taxi systems, public transport routing, and other VRP use cases.
The benchmark currently contains 144 instances
(16 cities × 9 variant combinations).
Supported Vehicle Routing Problem variants include:
- CVRP – Capacitated Vehicle Routing Problem
- VRPTW – Vehicle Routing Problem with Time Windows
- AVRP – Asymmetric Vehicle Routing Problem
- HVRP – Heterogeneous Vehicle Routing Problem
- MTVRP – Multi-Trip Vehicle Routing Problem
- TDVRP – Time-Dependent Vehicle Routing Problem
- SiDVRP – Site-Dependent Vehicle Routing Problem
- SDVRP – Split Delivery Vehicle Routing Problem
- VRPDB – Vehicle Routing Problem with Driver Breaks
The dataset provides the following variant combinations:
- CVRP + VRPTW + AVRP
- CVRP + VRPTW + AVRP + HVRP
- CVRP + VRPTW + AVRP + MTVRP
- CVRP + VRPTW + AVRP + TDVRP
- CVRP + VRPTW + AVRP + HVRP + MTVRP + TDVRP
- CVRP + VRPTW + AVRP + HVRP + SiDVRP
- CVRP + VRPTW + AVRP + SDVRP
- CVRP + VRPTW + AVRP + MTVRP + VRPDB
- CVRP + VRPTW + AVRP + HVRP + MTVRP + TDVRP + SiDVRP + SDVRP + VRPDB
Note:
The benchmark is designed to be flexible and application-driven, meaning users are not expected to evaluate all instances. Instead, it is recommended to select subsets of instances that best match the target real-world scenario (e.g., problem size, demand structure, and city topology). Some variants require other variants to be present (SiDVRP requires HVRP, VRPDB requires MTVRP).
Travel times are generated using OpenStreetMap data combined with a commercial routing API. More accurate travel-time estimation may require access to commercial traffic data sources.
The dataset includes instances generated for cities and regions of different sizes.
| City / Region | Instance Name | Type | Delivery Points | Supermarkets | Convenience | Supermarket TW | Convenience TW | Depot/Driver TW | Max Work (min) | Min Break (min) |
|---|---|---|---|---|---|---|---|---|---|---|
| Genoa | RW-GE75-VRP | City | 75 | 50 | 25 | 06:00–12:00 | 06:00–11:00 | 05:30–13:00 | 270 | 45 |
| Lyon | RW-LY85-VRP | City | 85 | 0 | 85 | — | 06:00–11:30 | 05:30–12:30 | 270 | 45 |
| Oklahoma City | RW-OK60-VRP | Metro | 60 | 20 | 40 | 06:00–15:00 | 06:00–18:00 | 05:30–19:00 | 480 | 30 |
| Central Dalmatia | RW-CD50-VRP | Region | 50 | 50 | 0 | 05:30–10:00 | — | 05:00–11:00 | 270 | 45 |
| City / Region | Instance Name | Type | Delivery Points | Supermarkets | Convenience | Supermarket TW | Convenience TW | Depot/Driver TW | Max Work (min) | Min Break (min) |
|---|---|---|---|---|---|---|---|---|---|---|
| Zagreb | RW-ZG170-VRP | City | 170 | 0 | 170 | — | 06:00–12:00 | 05:30–13:00 | 270 | 45 |
| Brasilia | RW-BR100-VRP | City | 100 | 100 | 0 | 06:00–14:00 | — | 05:30–15:00 | 330 | 30 |
| Copenhagen | RW-CP150-VRP | Metro | 150 | 100 | 50 | 05:30–11:00 | 05:30–10:30 | 05:00–12:00 | 270 | 45 |
| Ruhr Area | RW-RU120-VRP | Region | 120 | 40 | 80 | 05:00–12:30 | 05:00–14:00 | 04:30–15:00 | 270 | 45 |
| City / Region | Instance Name | Type | Delivery Points | Supermarkets | Convenience | Supermarket TW | Convenience TW | Depot/Driver TW | Max Work (min) | Min Break (min) |
|---|---|---|---|---|---|---|---|---|---|---|
| Singapore | RW-SG200-VRP | City | 200 | 200 | 0 | 06:00–13:30 | — | 05:30–14:30 | 300 | 60 |
| Melbourne | RW-ME340-VRP | City | 340 | 113 | 227 | 06:00–15:00 | 06:00–17:00 | 05:30–18:00 | 510 | 60 |
| Cairo | RW-CA240-VRP | Metro | 240 | 160 | 80 | 06:00–16:00 | 06:00–15:30 | 05:30–17:00 | 300 | 60 |
| SF Bay Area | RW-SF300-VRP | Region | 300 | 0 | 300 | — | 06:00–16:30 | 05:30–17:30 | 480 | 30 |
| City / Region | Instance Name | Type | Delivery Points | Supermarkets | Convenience | Supermarket TW | Convenience TW | Depot/Driver TW | Max Work (min) | Min Break (min) |
|---|---|---|---|---|---|---|---|---|---|---|
| London | RW-LO1000-VRP | City | 1000 | 0 | 1000 | — | 06:30–13:00 | 06:00–14:00 | 270 | 45 |
| Istanbul | RW-IS400-VRP | City | 400 | 400 | 0 | 05:00–13:00 | — | 04:30–14:00 | 270 | 45 |
| Mexico City | RW-MC800-VRP | Metro | 800 | 267 | 533 | 05:00–14:00 | 05:00–13:30 | 04:30–15:00 | 300 | 30 |
| Greater Bay Area | RW-GB600-VRP | Region | 600 | 400 | 200 | 06:00–15:30 | 06:00–18:00 | 05:30–19:00 | 240 | 20 |
When reporting results in scientific papers, please use instance names in the following format: RW--
Examples:
- RW-LO1000-VRP
Real-world London instance with 1000 delivery points for the base variant (CVRP + VRPTW + AVRP) - RW-LO1000-HVRP
Instance including HVRP (and 3 basic) constraints - RW-LO1000-HMTTDVRP
Instance including HVRP + MTVRP + TDVRP (and 3 basic) constraints - RW-LO1000-RWVRP Instance including all 9 variants (Real-World VRP) The abbreviations are ordered according to the following sequence: HVRP, MTVRP, TDVRP, SiDVRP, SDVRP, VRPDB.
This naming convention ensures consistency and comparability across different research works.
The repository is divided into two main folders:
-
src/ – contains the code used for dataset generation
-
data/ – contains all generated files, organized into 16 city folders
Within the src/ folder are scripts for dataset generation. Files required to define VRP instances are located directly in the city folders (data//).
Each city folder contains:
- data//locations.json – location data used for all VRP variants
- data//vehicles.json – vehicle data used for all VRP variants
- data//vehicle_location_accessibility.json – vehicle-to-location accessibility matrix used for SiDVRP
- data//time_matrix.json – static time matrix for non-TDVRP variants
- data//time_matrix_hvrp_8_palets.json – static time matrix for HVRP vehicles with 8-pallet capacity
- data//time_matrix_hvrp_18_palets.json – static time matrix for HVRP vehicles with 18-pallet capacity
- data//time_matrix_hvrp_26_palets.json – static time matrix for HVRP vehicles with 26-pallet capacity
- data//time_matrix_tdvrp.json – time-dependent matrix for TDVRP variants
- data//time_matrix_tdvrp_hvrp_8_palets.json – time-dependent matrix for TDVRP + HVRP variants with 8-pallet capacity
- data//time_matrix_tdvrp_hvrp_18_palets.json – time-dependent matrix for TDVRP + HVRP variants with 18-pallet capacity
- data//time_matrix_tdvrp_hvrp_26_palets.json – time-dependent matrix for TDVRP + HVRP variants with 26-pallet capacity
Additional files include:
- data//maps/ – plots used for validation and visualization
- data//solutions/ – stored solutions
- data//tests/ – additional test files
- data//tmp/ – temporary files
-
To reproduce the dataset, run
main.pyin its current state (you need to rungenerate_full_tdvrp_time_matrix(city),convert_3d_to_static(city)andgenerate_heterogeneous_matrices(city)). These functions generate all the required time matrices for different variants. -
Load the required files depending on your VRP variant:
data/<city>/locations.json– used for all VRP variants; SDVRP-specific demand and service time parameters are included when applicabledata/<city>/vehicles.json– used for all VRP variants; HVRP-specific vehicle capacity and VRPDB-specific driver parameters are included when applicabledata/<city>/vehicle_location_accessibility.json– used for SiDVRP variantsdata/<city>/time_matrix.json– static time matrix for non-TDVRP, non-HVRP variantsdata/<city>/time_matrix_hvrp_<capacity>_palets.json– static time matrix for HVRP variantsdata/<city>/time_matrix_tdvrp.json– time-dependent matrix for TDVRP variantsdata/<city>/time_matrix_tdvrp_hvrp_<capacity>_palets.json– time-dependent matrix for TDVRP + HVRP variants -
Generate a solution using your VRP algorithm and compare it with the best-known solution for the given instance.
A solution names its variant in instance_name and lists one route per used vehicle. Each route is a sequence of location IDs starting and ending at depot 0; intermediate depot visits are allowed only in MTVRP, and repeated customer visits only in SDVRP. In MTVRP, intermediate 0s mark depot returns between trips; they reset vehicle load, and every depot visit adds its configured service time. HVRP selects a travel matrix by vehicle capacity, TDVRP selects each leg's matrix by departure time, and SiDVRP routes must respect the vehicle-to-location accessibility matrix.
For SDVRP, demands is required and aligns with route: depot entries are 0, and customer entries are the quantities delivered on those visits. Split quantities across visits must sum to the customer's demand_sdvrp. Unloading is calculated per visit as 5 minutes plus 120 seconds per unit at supermarkets or 90 seconds per unit at convenience stores; each split visit incurs the 5-minute base time. For VRPDB, breaks_after contains route indexes where a break is taken after the stop and before the next leg, subject to the vehicle's driving limit.
For VRPDB, breaks_after uses zero-based indexes into the route. Validation reports total_time_consumed (driving, waiting, service, and breaks), driving_time (travel only), and cost. Cost is driving time plus 10,000 per used vehicle for variants that include MTVRP.
A more detailed README for solutions, including a small example of an input and its corresponding solution, is available under data/DummyCity. The examples can be run from main.py using the validate_solutions function. This function can also be used by users to validate their own solutions for any city in the benchmark, allowing them to check whether a solution is valid and inspect the resulting validation metrics.
Planned future work includes storing the best-known solutions.