How STARK VRP Cuts Chinese Logistics Costs with AI‑Powered Routing

With China's logistics sector accounting for 18% of GDP, vehicle routing is a critical cost factor; reducing the number of trucks and travel distance directly lowers logistics expenses.

Cainiao's AI department, together with IDST and Alibaba Cloud, created STARK VRP, a distributed vehicle‑routing solver that meets complex Chinese business needs while achieving near‑international performance.

VRP Types Supported by STARK

CVRP – capacity constraints (volume, weight, customers, max distance)

VRPTW – time‑window constraints

VRPPD – pickup and delivery (e.g., O2O food delivery)

MDVRP – multiple depots

OVRP – open routes (no return to depot)

VRPB – backhaul

Heterogeneous Fleet – mixed vehicle types

T+n – flexible delivery dates to merge orders

Milk Run – cyclic pickup

Skilled VRP – driver‑customer matching

Same Route VRP – orders on a single route

Generalized VRP – multiple possible pickup locations

Split Delivery – split orders across trucks

VRP with intermediate facilities – charging stations for EVs

2E VRP – multi‑stage transport (truck to van or drone)

Algorithmic Framework

Traditional exact solvers cannot handle large‑scale data; STARK builds a meta‑heuristic framework that includes Large Neighborhood Search, Adaptive Large Neighborhood Search, Variable Neighborhood Search, Metaheuristic Hybrids, Iterated Local Search, Memetic Algorithm, Tabu Search, Simulated Annealing, Guided Local Search, and Fast Local Search.

ALNS – Adaptive Large Neighborhood Search

ALNS uses a bandit‑based hyper‑heuristic to select operators dynamically, adapting to problem stages much like choosing cavalry or tanks in different terrains.

Parallelization (ISLAND & EE Pool)

The ISLAND model exchanges solutions among islands to maintain diversity and avoid local minima. EE Pool balances exploration and exploitation across islands without global synchronization, using a stochastic sampling process (SSP).

Deep Reinforcement Learning

A reinforcement‑learning layer embeds problem instances and updates operator selection probabilities based on feedback, aiming for data‑driven improvements in efficiency and solution quality.

Business Impact

In the Villager (CunTao) business, STARK reduced travel distance by 28%; in Retail‑Tong, vehicle count dropped 10% and distance by ~13% (3000 km → 2600 km). The solver matches several Best‑Known Solutions on benchmark instances (400‑job and 1000‑job tests).

Conclusion

STARK VRP demonstrates that AI‑enhanced vehicle‑routing is essential for modern logistics, complementing autonomous driving and serving both internal Alibaba services and external partners such as Ruirun and YunNiao.