Last-mile delivery systems operate in a world that rarely stands still. Orders arrive while trucks are already moving, traffic conditions shift, and delivery windows are often tight. In this environment, a static plan—a route calculated once at the start of the day—is rarely enough. Instead, dispatchers must constantly update their plans, inserting new stops into existing routes or rearranging them when a customer cancels or a vehicle breaks down. This constant need to re-optimize creates a difficult computational challenge: the system must find a new, valid arrangement of stops in a matter of seconds. If the computer takes too long to think, the delivery window closes, or the driver waits idly. If the computer guesses a new route that violates the rules—such as a truck carrying too much weight or arriving too late—the entire plan collapses, and the system must start over. The goal is not just to find the shortest path, but to find a workable path, instantly.

Researchers at Hubei University of Automotive Technology have developed a new approach to this problem, one that borrows a strategy from the world of artificial intelligence to make these split-second decisions more reliable. They call their system SRI-ECR-balanced. Rather than trying to calculate the perfect route from scratch every time a new order arrives, the system works like a skilled editor who drafts a few sentences, checks them for errors, keeps the good parts, and immediately fixes the bad ones. This method, known as speculative execution, allows the system to propose several future stops at once, verify which ones are safe, and discard the rest before they cause a problem. By doing this, the system can recover from conflicts much faster than traditional methods, ensuring that a delivery plan remains feasible even when the situation on the ground changes rapidly.

The core of this new framework is a cycle of drafting, verifying, and repairing. When a new order comes in, a "draft generator" proposes a short sequence of customers to add to a truck's route. In the past, systems might have tried to insert just one customer at a time, which is slow, or they might have tried to insert a long list of customers all at once, which often leads to mistakes. This new system strikes a balance: it proposes a small group of stops, perhaps five at a time. A separate "verifier" then checks this group against the rules of the road, such as the truck's weight limit and the customer's time window. The system accepts the longest stretch of the proposed group that fits the rules and immediately discards the rest. If the first three stops in the proposed group are fine but the fourth is too heavy for the truck, the system keeps the first three and rolls back the fourth. It does not throw away the whole plan; it only removes the part that broke the rules.

Once the invalid stops are rolled back, a specialized "repair expert" steps in to fix the problem. This is not a generic fixer; the system diagnoses exactly what went wrong. If the issue is that the truck is too full, the expert prioritizes finding a different truck with more space. If the issue is that a customer will be missed because of a tight time window, the expert looks for a route that can arrive earlier. This targeted approach is crucial. The researchers found that without this specific, conflict-aware repair step, the system failed to find any valid routes in stressful situations. When the system was forced to use a random repair method instead of an expert one, it could only fix about one-third of the problems, leaving many orders unserved. With the expert router, the system successfully repaired every single conflict in their tests, maintaining a perfect record of feasible plans.

To make this process fast enough for real-time use, the system also uses a form of memory called a route-prefix cache. In a dynamic environment, most of a truck's route remains the same from one moment to the next; only the end of the route changes. Traditional systems often re-calculate the entire journey every time they make a small change, wasting time on parts of the trip that haven't changed. This new system remembers the calculations for the parts of the route that are already confirmed. When it needs to check a new stop, it only calculates the new part, reusing the stored data for the rest. This simple trick made the system more than three times faster at evaluating potential routes. In tests on a standard set of 100 customers, this caching allowed the system to evaluate routes 3.12 times faster than if it had to start from scratch. On a larger test with 200 customers, the speedup was 3.08 times.

The researchers tested their system against existing tools used by logistics companies, including OR-Tools and PyVRP, which are powerful solvers designed to find the absolute shortest routes. They ran simulations where new orders arrived constantly and where time windows were made artificially tight to create stress. In these difficult conditions, the traditional solvers took nearly the full three seconds allowed for each decision to find a solution, and even then, they sometimes produced routes that were slightly shorter but took longer to compute. The new system, by contrast, found a valid solution in less than half a second on average. While the routes it produced were about 7.8% longer in total distance than those found by the powerful solvers, the trade-off was deliberate. In a real-world delivery scenario, a slightly longer route that is ready instantly is often more valuable than a perfect route that arrives too late. The system prioritized getting a workable plan out the door over squeezing out every last mile of efficiency.

The study also explored how the system behaves under different levels of risk. They compared a "balanced" approach, which carefully weighs the risk of adding a new stop, against an "aggressive" approach that tries to add as many stops as possible before checking. The aggressive method caused the system to make many mistakes that had to be rolled back, slowing it down significantly. The balanced approach, which penalizes risky moves before they are even proposed, reduced the number of mistakes and kept the system moving smoothly. This confirmed that the system works best when it speculates with caution, proposing a few steps ahead but checking them rigorously before committing.

The researchers did not claim that their system is the best at finding the shortest possible route in a static world where time is not a factor. They explicitly stated that for problems where the only goal is to minimize distance and the computer has several seconds to think, the traditional solvers still win. The value of this new framework lies specifically in the chaotic, time-pressured world of dynamic dispatch. In these scenarios, the ability to recover from a conflict instantly and maintain a valid plan is more important than a marginal improvement in distance. The system proved its worth by serving 100% of customers in the most stressful test scenarios, with zero rejected orders, while keeping the response time low enough to be useful for a live dispatch center.

Even when the researchers introduced random noise to simulate unpredictable traffic delays, the system held up. It continued to serve all customers and repair conflicts successfully, showing that the method is robust enough to handle real-world uncertainty. They also tested the system on a slice of real delivery data from a major logistics company, using actual drop-off coordinates. The system ran successfully on this real-world data, serving all 80 stops in the test set. However, the researchers were careful to note the limits of their current design. When they tested a much larger scenario with 400 customers, the system could not find a perfect solution within the time limit, serving only about three-quarters of the orders. This indicates that while the method works well for medium-sized fleets, larger operations would require additional strategies to break the problem down into smaller pieces.

Ultimately, this work demonstrates a shift in how we think about solving routing problems. Instead of viewing the computer as a calculator that must find the single best answer, the researchers treated it as a decision-maker that must find a good answer quickly. By adopting the speculative patterns used in modern artificial intelligence—drafting, verifying, and caching—the system can navigate the uncertainty of dynamic delivery with speed and reliability. The findings suggest that for many real-world logistics challenges, the ability to adapt instantly is the true measure of a successful system, even if it means the routes are not mathematically perfect. The system does not promise to eliminate all delays or find the shortest path every time, but it does promise that when a new order arrives, a plan will be ready almost immediately, keeping the wheels of the delivery network turning.