Learning from the past to invent the future of last-mile logistics

From the September-October 2024 edition of Supply Chain Management Review.

September-October 2024

Back in late 2023, in response to global panic about the state of the supply chain, President Joe Biden announced the formation of the White House Council on Supply Chain Resilience. “We’re doubling down on our…

Browse this issue archive.

Access your online digital edition.

Download a PDF file of the September-October 2024 issue.

The meteoric rise of e-commerce has transformed last-mile supply chains as companies strive to stay ahead of changing consumer demands and technological innovation. Technologies—notably artificial intelligence and machine learning—are fueling another phase in this evolution that is bringing more transformative change.

It is impossible to know exactly what the last mile will look like a decade from now, but the outlines are emerging. Key areas such as vehicle routing and order fulfillment will undergo enormous change to meet ever-increasing customer expectations and efficiency demands, and the need to flex with uncertain market conditions.

Tools upgrade needed

Companies have made huge strides over the last decade or so in revising their supply chain configurations to enable the development of high-performance last-mile operations. However, to keep pace with e-commerce’s next evolutionary phase, the industry must take these efforts to a higher level.

Consider, for example, vehicle routing. Many decades of research and technical know-how have focused on the problem of how to route delivery vehicles that make multiple stops most efficiently. Today’s routing algorithms and models effectively capture the real-world complexity of last-mile operations.

However, these tools are not good enough to achieve the performance levels required to meet future e-commerce demands. These shortcomings have become unacceptable as the pressure to customize and speed up last-mile deliveries has increased. And this trend will continue for the foreseeable future. We must extend the boundaries that define what vehicle routing methods can do. For example, the same family of methods used to power AI-enabled large language models, such as ChatGPT, also offers huge potential for solving complex optimization problems in the logistics industry such as the vehicle routing problem.

Similarly, advances in predictive analytics over recent decades are not enough to meet the needs of future last-mile operations.

For example, demand forecasting has improved markedly, and practitioners now have very capable methods for forecasting demand. However, these methods need to be more precise. The ability to predict how often a product will be bought next week is no longer sufficient. Because of the increasing demand for high-speed e-commerce services, companies will want to know how many items customers will purchase in specific zip codes within the next 30 minutes. Traditional methods that rely on high volumes of data to make such estimates are incapable of achieving this level of precision.

Research makeover

The industry also needs new research approaches to help it keep pace with fast-changing online markets. Recent changes to research programs at the MIT Center for Transportation & Logistics (MIT CTL) exemplify the required  shift in emphasis.

The MIT Megacity Logistics Lab came into existence about 10 years ago with a focus on researching the logistics of moving freight in large cities. This mission made sense because the transportation infrastructure in these long-established population centers was ill-equipped to handle modern-day challenges such as traffic generated by e-commerce.

This is still the case today although we have developed ways to ease traffic-related problems. However, when the lab was founded, there was much less recognition of how important last-mile logistics would become, let alone the inception of more demanding services like two-hour delivery models.

Also, operations research or optimization methods broadly defined the lab’s approach. This research approach typically involves using mathematical models to formulate problems in areas such as vehicle routing. Mathematical equations represent the problems, and algorithms are used to solve the equations. However, when applied in the real world—for example, to optimize a delivery route with 120 stops—it can take a long time to find the optimal solution, or such a solution may not be forthcoming.

Heuristic approaches are another family of methods that fall under the traditional operations research banner. Basically, heuristics comprise a set of rules or decision procedures that algorithms tailored to a particular problem can incorporate. For example, encoding rules that define a good move in a routing problem. These methods tend to be faster and better at finding solutions, but there is no guarantee that the solutions are optimal.

Nowadays, industry developments are happening faster than these traditional methods can keep up. For instance, if a company needs to optimize the logistics system that supports on-demand, one-hour delivery services, there is little time to develop optimized routes. Also, the trend toward customer-centricity introduces more uncertainty to logistics systems, which traditional methods struggle to capture adequately.

These advanced routing solutions are also more adept at adjusting route plans in response to changing conditions. Soon, drivers will use voice recognition technology to update systems instantly about real-time problems such as adverse weather, changes in local parking regulations, or unexpected road closures. This information will immediately update and refine route planning models for delivery fleets. 

Today, we need research methodologies that are fast, more data-driven, and excel at learning from real-world operational data. Such methods are prevalent in programs driven by AI and machine learning.

To meet demands like these, MIT CTL launched the Intelligent Logistics Systems Lab (ILS) in June 2024 in collaboration with intralogistics company Mecalux. The lab is pursuing research at the intersection of operations research, AI, and machine learning technologies in multiple areas, including predictive and prescriptive analytics, autonomous logistics systems, and human decision-making enhanced by combining human intelligence with AI.

Shape of e-commerce to come

Research like this paves the way for last-mile logistics systems aligned with the future of e-commerce. An example is the development of advanced dynamic route planning solutions. AI-inspired algorithms can absorb lessons from historical information on driver behavior and location patterns and enrich these insights with data from external sources, such as public road networks and weather databases. Instead of only finding the shortest routes to customer addresses, next-generation planning systems can craft routes optimized for safety, sustainability, driver ergonomics, customer satisfaction, and minimal delays.

These advanced routing solutions are also more adept at adjusting route plans in response to changing conditions. Soon, drivers will use voice recognition technology to update systems instantly about real-time problems such as adverse weather, changes in local parking regulations, or unexpected road closures. This information will immediately update and refine route planning models for delivery fleets. 

AI will also use drivers’ insights to predict optimal delivery windows based on experience. A driver might note that a customer prefers items to be left in a certain area at the designated address. The AI system integrates the information and tailors future deliveries to this preference, enhancing efficiency and customer satisfaction. Algorithms will also determine when individual buyers are available to receive packages.

Future last-mile technology could also bring AI-driven micro-hubs, scattered across urban landscapes, that fulfill immediate delivery requests.

The benefits of AI-powered delivery go beyond the interface with consumers. The development of AI models that generate hyper-local forecasts of near-term demand with high accuracy is another important area of research, enabling companies to preemptively position inventory closer to customers. In addition to supporting faster deliveries, AI-informed inventory localization will reduce excess inventory across a supply network, reduce vehicle mileage, and improve the environmental footprint of the entire delivery process.

The future is hybrid

Looking further into the future, if wide-scale autonomous delivery services become a reality, they should benefit from the tactic knowledge accumulated by traditional services. For instance, knowledge of customer idiosyncrasies or local parking restrictions would be part of trained models running in the background that inform autonomous vehicles on how best to complete deliveries.

This combination of traditional and new methods will be a critically important part of future last-mile systems.

The ILS was born out of a belief that the work required to develop a new family of e-commerce logistics solutions should combine the best of traditional and new worlds. Past research shows that the greatest potential for improvement does not necessarily lie in replacing the methods used over the last one or two decades, but in using AI and machine learning to advance them.

The author provides a more detailed account of future last-mile developments and the Intelligent Logistics Systems Lab’s research agenda in the Vehicle Routing in the Age of AI podcast. To hear the podcast go to ctl.mit.edu/podcasts/vehicle-routing-age-ai

SC
MR