Order picking remains one of the biggest cost drivers in warehouse logistics. The pressure is particularly acute in e-commerce, where single-item or small-basket orders are assembled from multiple storage locations across a facility.
At the FM Logistic site involved in the project, picking spans more than 17,700 pick locations. Operators use batch picking, collecting items for multiple orders in one run through the warehouse. That makes order grouping critical: the better the batches are built, the fewer trips are needed and the shorter each route becomes. At this scale, even small batching errors can translate into thousands of extra kilometres walked or driven each year, as well as higher wear on warehouse vehicles.
FM Logistic says its process was already highly tuned, but it still believed there were incremental gains to capture. Traditional optimisation methods, however, were starting to reach their limits.
From an in-house algorithm to a DeepMind pilot
The work started before AlphaEvolve entered the picture. FM Logistic first built its own order-grouping algorithm, which improved picking performance. From there, the focus shifted from “fixing the basics” to finding the next few percentage points in a process that was already structured and optimised.
The company was then invited to test Google AlphaEvolve, a solution developed by Google DeepMind that uses Gemini models to generate and refine algorithms. FM Logistic says it was among the first organisations worldwide to join the early-access programme.
AlphaEvolve is designed to tackle complex optimisation problems by proposing new algorithm variants, testing them and scoring the results.
For FM Logistic, the objective was to find a better way to build picking batches. Each suggested change was validated in a dedicated simulation environment intended to mirror real warehouse operating conditions.
The make-or-break step: defining how to score “better”
FM Logistic says a key part of the project was building an evaluation model for the options AlphaEvolve produced. Simply shortening routes could not be the only measure of success.
The company notes that AI can quickly surface improvements that are not obvious — but not every “best” result is suitable for a live warehouse. The team therefore defined process rules, constraints and parameters in detail. FM Logistic says combining operational know-how with AlphaEvolve’s capabilities helped it arrive at solutions that were both effective and practical to implement.
Results in day-to-day operations
The algorithm ultimately deployed delivered an improvement of more than 10% compared with the previous approach, FM Logistic says. In operational terms, that meant shorter picking routes and a reduction in annual operator travel of more than 15,000 kilometres.
FM Logistic adds that the outcome is notable given the starting point: the process had already been optimised multiple times, and the company’s in-house solutions had significantly reduced picking-route length. Even so, it says AlphaEvolve helped identify additional improvements in an area where each new gain is increasingly difficult to achieve.
The project also underlined a broader point, the company argues: productivity gains do not always require new automation or facility upgrades. In this case, the benefits came from using data more effectively and applying a more advanced approach to process optimisation.
FM Logistic says the gains were achieved without investment in new infrastructure or equipment, relying instead on data and process design.
