中EN
Back to the archive

The encyclopedia · Engineering & Operations · Technical decision · 2016–2021

Alibaba taught machines to route a city's delivery in seconds, not in planners' heads

Cainiao built ALNS + reinforcement-learning vehicle-routing algorithms to turn half-hour delivery promises into dispatchable routes.

Alibaba Group · Cainiao Network

the move

Alibaba pioneered integrated online and offline retail, so an order placed on a phone can be delivered in as little as 30 minutes from a nearby store. Fulfilling that promise for millions of orders a day is a routing problem no human desk can solve.

Its supply-chain and operations-research group built a vehicle-routing (VRP) algorithm suite combining Adaptive Large Neighborhood Search (ALNS) with deep reinforcement learning. It reaches roughly 80-90 percent of optimal performance in a fraction of the time of standard methods.

The system is used across Hema, Cainiao and Lazada: guiding how items are placed in a warehouse so pickers move efficiently, optimizing multi-party transport, and returning a real-time route in milliseconds for last-mile delivery.

why it works

  • It attacked routing at the scale and speed the business actually needed, not at academic pace.
  • Combining ALNS with deep reinforcement learning made it fast enough to re-route in real time.
  • Algorithm developers were embedded in the operation, so solutions were built and shipped continuously.
  • It works across picking, line-haul and last mile, so the same engine covers the whole chain.
the payoffPut the solver inside the business, not a consultantinspired

what transfers

Handling an operations problem that arrives at machine speed means owning the solver and running it inside the business — not commissioning a one-time optimization study.

what came after

The work was named a finalist for the 2021 INFORMS Franz Edelman Award, recognizing it as one of the most advanced applied operations-research projects in the world, and the algorithms ran across Alibaba's major logistics platforms. The team, led by long-time OR researchers embedded in the business, framed it as proof that algorithm developers should be part of operations, not external consultants.

references

spotted an error? The archive wants to know.

same kind of clever

2023 · Engineering & Operations

Kuaishou trained its recommender to keep you coming back, not just to get a click.

Reward the gap between sessions, not just the item you show
2019–2022 · Engineering & Operations

Deep reinforcement learning cut Lenovo laptop scheduling from six hours to 30 minutes.

Train the scheduler itself instead of hand-writing rules
2017–2021 · Engineering & Operations

JD.com ran its robot warehouses on dispatching algorithms, not more pickers

Dispatch the robots, not the people who pick by hand