Smarter algorithms stop factory robots from colliding

How researchers helped JD.com improve the efficiency of its fulfillment warehouses

Credit: Sebastian Thibault

Michael Maiello | Jun 21, 2021

Sections Strategy

Collections Operations Management

China’s e-commerce warehouses shipped more than 83 billion parcels in 2020, a nearly 23-fold increase from a decade ago. This online-shopping boom creates massive logistics problems for retailers such as JD.com, which sells everything from books to appliances, offering same-day delivery to 470 million customers in China who place orders by 11 a.m., whether they buy something big or small, common or hard to get. 

The solution for this enormous logistics challenge could be robot-driven automation, according to a team of researchers from JD.com, Shanghai University of Finance and Economics, Chicago Booth, and the University of Southern California. In work they conducted at JD.com, the researchers helped the company improve the efficiency of its fulfillment warehouses by adopting a model of facility management in which workers stay at stations while robots bring them products to be packed. The findings could one day help companies deliver products efficiently by drone and robots, they write. 

JD.com is comparable to the better-known Alibaba in terms of revenue ($114 billion in 2020 for JD.com compared with $109 billion for Alibaba). But JD.com manages its own inventories and fulfillment while Alibaba operates more as an online marketplace, akin to eBay or Etsy. To get goods into boxes and onto trucks and planes, JD.com has largely automated its warehouses and handles its own shipping. (Imagine if Amazon owned FedEx.) 

In its use of robots, JD.com flips the conventional human-driven fulfillment-center model. In a standard warehouse, people rush around on foot or forklift, grabbing and packing products as orders come in. This is called a picker-to-parts model. JD.com uses a parts-to-picker system, where swarms of football-size robots scurry across fulfillment-center floors to bring racks of materials from shelves to humans waiting to grab and package products. The system increases productivity, preserves worker safety, and cuts costs, the researchers write.

However, to make parts-to-picker work, each of the warehouses’ hundreds of robots must solve the complex problem of which shelves to bring to which packers once every five seconds—without colliding with each other. JD.com initially tried to accomplish this by using off-the-shelf artificial-intelligence software to control the robots, but “commercial software proved too slow,” says Chicago Booth’s Linwei Xin, one of the researchers involved. Because of this, the company found that many of its automated warehouses didn’t ship enough packages to justify the construction costs. 

About five years ago, the researchers began developing algorithms to improve the functioning of the warehouse robots. They broke the problem into discrete layers, or challenges. For example, an integrated management layer meshes fulfillment with other systems such as order processing. Each layer is defined by multiple algorithms, and all are continuously monitored and updated to help the automated centers deal with changing demand and conditions. 

The algorithms the researchers created allow the company’s warehouses to function smoothly even when orders accelerate to 10 times normal, as they did during the worst of the coronavirus pandemic, according to the research. While many conventional factories slowed operations or were shut entirely to prevent the spread of the virus during the height of the recent pandemic, JD.com’s warehouses were able to operate thanks to the algorithm-driven parts-to-picker system.

JD.com recorded a decrease in its fulfillment expense ratio (which measures fulfillment costs to sales) to 6.5 percent in 2020 from 7.2 percent in 2016. Also, in 2020, 90 percent of orders sold by JD.com, rather than third-party sellers, were delivered on the same day or the day after they were placed, the researchers report. 

Xin sees potential outside the warehouse, perhaps to control drone or robot delivery of products to urban areas, or to deliver items such as vital medical supplies to hard-to-reach areas. 

However, “100% automation is not ideal,” the researchers write. Even the most automated warehouses still need people to conduct oversight, service the robots, and pack nonstandard items, they note. That said, they also have yet to find a point at which investing in robots yields diminishing benefits.