1st-hand & in-depth info about Alibaba's tech innovation in AI, Big Data, & Computer Engineering
How AI-powered warehouse is transforming the logistics industry
What do warehousing and computing have in common? As Alibaba logistics affiliate Cainiao’s Zhu Lijun told audiences at the 2018 Global Smart Logistics Summit in Hangzhou, their common reliance on storage, extraction, and processing lends the two some striking operational and structural parallels. For all their similarities, though, linking these physical and digital realms for smarter, scalable distribution solutions challenged developers at Cainiao to adapt new approaches en route to their latest milestone for the online commerce sector: a completely AI-powered warehouse founded on flexible automation technology.
In Hangzhou, Mr. Zhu presented just that achievement, taking the stage alongside a model of Cainiao’s smart warehouse complete with package-tending robots stationed at conveyor belts. Along the way, his team redeployed from its usual quarters to do research and coding on-site at Cainiao’s warehouses, incorporating firsthand insights while studying how facilities could be restructured to better suit automation. By synthesizing capabilities spanning automated guided vehicles (AGVs), 3D camera sensing, robotic arm manipulation, and ideas borrowed from MapReduce technologies in big data processing, the team successfully advanced a “swarm intelligence” AI model in which a large number of robots work collaboratively while bidding to make optimal use of warehouse resources.
Drawing on highlights from Mr. Zhu’s speech, we look today at the technical and conceptual challenges Cainiao overcame to launch an autonomous logistics solution set to solve pain points in the labor market and engage rising consumer demand for rapid delivery.
By official estimates, China’s logistics labor force will fall by 25% before 2050, a shift already underway with the rapid rise of e-commerce and early achievements in automation. The projected decline also reflects changes beyond human obsolescence, with recruitment suffering from a shortage of qualified candidates in the economically crucial southeastern coastal regions. Adding to the burden for groups like Cainiao, new hires frequently lag behind experienced workers by as much as 50% in efficiency, while their learning curve continues to steepen with the increasing diversity of goods and orders introduced by the development of e-commerce. With consumers now taking near-immediate delivery for granted, human labor may not be adequate even to maintain present logistical efficiency, let alone improve on current standards.
In Alibaba’s founder Jack Ma’s words, “Cainiao must do its utmost to build a national smart logistics backbone network to reduce the proportion of China’s logistics costs below 5% of GDP.” As Zhu Lijun conveyed, doing that utmost now means seeking solutions that not only support automation but further refine its scalability, a task requiring the wedding of two areas in innovation that have so far remained largely separate: AI and automated logistics, known together as flexible automation.
Flexible automation outpaces traditional automation in two key regards. First, it is highly scalable and agile, making it largely impervious to systemic failure. As well as quickly deploying robots and automated equipment to deal with rising order counts, this model can quickly disengage processes from failures at any single point, allowing system-wide operations to continue while the error is corrected. Traditional automation, by contrast, will tend to suffer continuously from hindrances in processes until they can be resolved.
Second, flexible automation boasts a modular design, allowing facilities to prioritize shifting demands for the diverse range of goods stored on site while coordinating warehouse activities accordingly. Whether these fluctuations are seasonal or momentary, the machines do not rely on commands issued by human supervisors to adjust their task flow, instead, acting on priorities determined by computers that interpret incoming orders alongside inventory data.
Adding to the innovative promise of this approach, the development AI enables in logistical facilities in turn furnishes an evolving application setting for further developing AI capabilities.
Building a fully operational unmanned warehouse challenged Cainiao to automate every link of order fulfillment, spanning picking, packing, stacking, loading, and outbound to the distribution center and then, delivery site. To do so, the development team integrated its AI systems with several mature industry technologies ranging from robotic hardware to big data software.
To move items from one part of the warehouse to another, Cainiao deployed the same core technology used in conventional warehousing: AGVs. Whereas conventional AGV applications require periodic input or manual operation from human agents, Cainiao’s warehouse AGVs are controlled by a central job scheduling and dispatching system. In this way, the AI that handles picking processes can continuously redirect a fleet of more than 500 AGVs to make optimal use of each vehicle under changing circumstances.
To deal with the sheer volume and variety of goods processed in warehouses, Cainiao adapted big data framework MapReduce to support a modular, multi-zone parallel automation solution. MapReduce supports the coordination of picking activities in multiple zones designated for specific goods and the modes of operation needed to handle those commodities, later merging orders at a rebin zone. The most important factor in efficient automation for a modular layout is path planning for the AGVs that navigate these zones, especially to prevent traffic congestion or deadlock. To coordinate their fluent operation, sophisticated multi-robot pathfinding algorithms are used to monitor the collective compatibility of AGVs’ assignments while reassigning them to deal with incoming orders as circumstances change.
The last component needed to fully replace manual labor in an unmanned setting was a system of robotic arms that could interact with the AGVs. Cainiao applied advanced 3D camera and computer vision algorithms for object and position recognition to plan the movement of the robotic arms, allowing them to effectively segment packages and deliver them to shelves, conveyor belts, and AGVs throughout the warehouse.
Following extensive adaptation, these established technologies enabled Cainiao to meet industrial production line standards while achieving higher efficiency and reliability, as well as stronger resilience throughout the system to reduce the impact of malfunctions in specific processes.
Because flexible automation synthesizes hardware and software system engineering, implementing it requires thinking beyond categorical boundaries to incorporate a range of technologies into discrete processes. Along with the previously discussed robotics and big data capabilities, Cainiao’s unmanned operations reflect advances in the internet of things and edge computing technology, resource allocation and traffic control AI, and the emerging AI phenomenon known as swarm intelligence.
To explain the significance of swarm intelligence, Zhu Lijun presented two alternative frameworks for resource allocation which Cainiao can deploy depending on the size of a warehouse: centralized planning, and distributed bidding. In either case, resources refer to shelf space, goods in stock, and AGVs, with the goal of allocating these for maximum output in a given span of time.
For Cainiao’s smaller warehouses, centralized planning will tend to provide an effective approach to organizing autonomous machinery. Not to be confused with human supervision, centralized planning uses algorithms to calculate the most expedient use of available resources. The AI, in this case, processes information on the whereabouts and availability of AGVs and storage areas in the warehouse, then issues a stream of instructions to direct traffic while continuing to process new information.
To orchestrate flexible automation for larger warehouses, Cainiao adopted an organizing principle called distributed bidding, bringing its operations closer to a swarm intelligence model in which AGVs behave like fish in a school or birds in a flock. Treating AGVs as individual agents with specific availabilities to complete picking tasks, distributed bidding allows the central AI to award work to the most efficient candidates as they bid for incoming tasks. While this technology remains in an interim state of development, Cainiao believes that in the future each agent will have sufficient navigation and decision-making capabilities to perform such a role autonomously, volunteering itself where needed rather than answering commands from a central AI mechanism.
Highly scalable, easy to deploy, and modular — these attributes make flexible automation far more than a familiar replacement for human labor introduced to cut operating costs. As e-commerce grows to present order fulfillment demands beyond the scope of traditional warehousing, flexible automation offers a means not only to deploy helpful robotics, but also to reconceive of what robotics can achieve in practice.
As impressive as these developments are, Zhu Lijun ended his presentation by saying they are only the beginning of a coming revolution in logistics. Beyond e-commerce and order fulfillment, Cainiao’s work has broken new ground in the larger development of AI for physical systems, offering an early demonstration of how swarm intelligence models may soon come to organize autonomous functionality in clusters of robots and devices. Just as remarkably, it has adapted IT and computing to work autonomously in highly physical systems, a challenging leap even where human supervision is allowed to remain a factor in processes.
Looking forward, Cainiao aims to continue developing AI solutions for its logistics operations in association with Alibaba, and hopes its achievements will enable a positive shift in focus for human labor in the logistics industry.
(Original talk by Zhu Lijun朱礼君)