---
license: mit
language:
- en
tags:
- Optimization
- combinatorics
- Vehicle Routing Problem
- SVRP
- Logistics
- Transportation
pretty_name: SVRPBench Dataset
size_categories:
- 10K<n<100K
---
# 🚚 SVRPBench

SVRPBench is an open and extensible benchmark for the Stochastic Vehicle Routing Problem (SVRP). It includes 500+ instances spanning small to large scales (10–1000 customers), designed to evaluate algorithms under realistic urban logistics conditions with uncertainty and operational constraints.

## 📌 Overview

Existing SVRP benchmarks often assume simplified, static environments, ignoring core elements of real-world routing such as time-dependent travel delays, uncertain customer availability, and dynamic disruptions. Our benchmark addresses these limitations by simulating urban logistics conditions with high fidelity:

- Travel times vary based on time-of-day traffic patterns, log-normally distributed delays, and probabilistic accident occurrences
- Customer time windows are sampled differently for residential and commercial clients using empirically grounded temporal distributions
- A systematic dataset generation pipeline that produces diverse, constraint-rich instances including multi-depot, multi-vehicle, and capacity-constrained scenarios

## 📦 Dataset Components

The dataset includes various problem instances:
- Problem sizes: 10, 20, 50, 100, 200, 500, 1000 customers
- Variants: CVRP (Capacitated VRP), TWCVRP (Time Window Constrained VRP)
- Configurations: Single/Multi-depot, Single/Multi-vehicle

Each instance includes:
- Customer locations
- Demand volumes
- Time window constraints
- Vehicle capacity limits
- Depot coordinates

## 🧪 Supported Algorithms

The benchmark includes implementations of several algorithms:
- OR-tools (Google's Operations Research tools)
- ACO (Ant Colony Optimization)
- Tabu Search
- Nearest Neighbor with 2-opt local search
- Reinforcement Learning models

## 📊 Benchmarking Results

Results compare algorithm performance across different problem sizes:

| Model    | CVRP10 | CVRP20 | CVRP50 | CVRP100 | CVRP200 | CVRP500 | CVRP1000 |
|----------|--------|--------|--------|---------|---------|---------|----------|
| OR-tools | 1.4284 | 1.6624 | 1.3793 | 1.1513  | 1.0466  | 0.8642  | -        |
| ACO      | 1.5763 | 1.7843 | 1.5120 | 1.2998  | 1.1752  | 1.0371  | 0.9254   |
| Tabu     | 1.4981 | 1.7102 | 1.4578 | 1.2214  | 1.1032  | 0.9723  | 0.8735   |
| NN+2opt  | 1.6832 | 1.8976 | 1.6283 | 1.3844  | 1.2627  | 1.1247  | 1.0123   |

## 🛠️ Usage

```python
# Example of loading a dataset
from datasets import load_dataset
ds = load_dataset("MBZUAI/svrp-bench", split="test")
ds[0]
```

### Sample
```json
{'appear_times': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 'demands': [0, 33, 52, 35, 85, 77, 68, 17, 61, 32, 23],
 'file_name': 'cvrp_10_multi_depot_multi_vehicule_capacities.npz',
 'instance_id': 0,
 'locations': [[523, 497],
               [394, 344],
               [536, 599],
               [341, 412],
               [734, 652],
               [492, 569],
               [491, 238],
               [419, 787],
               [688, 422],
               [708, 490],
               [431, 454]],
 'num_vehicles': 13,
 'subset_name': 'cvrp_10_multi_depot_multi_vehicule_capacities',
 'vehicle_capacities': [40.0]}
```

## 🔑 Features

- Comprehensive evaluation framework for VRP algorithms
- Realistic travel time modeling with time-dependent patterns
- Time window constraints based on empirical distributions
- Support for multi-depot and multi-vehicle scenarios
- Visualization tools for solution analysis
- Extensible architecture for adding new algorithms

## 📚 Citation

If you use this benchmark in your research, please cite:

```bibtex
@misc{svrbench2025,
  author = {Heakl, Ahmed and Shaaban, Yahia Salaheldin and Takáč, Martin and Lahlou, Salem and Iklassov, Zangir},
  title = {SVRPBench: A Benchmark for Stochastic Vehicle Routing Problems},
  year = {2025},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/yehias21/vrp-benchmarks}}
}
```

## 📄 License

This project is licensed under the MIT License.


- 📄 Paper: [Arxiv](https://arxiv.org/abs/2505.21887)