Warehouse routing: −10.4%

FM Logistic, a global logistics provider operating in over 14 countries, ran the Poland pilot in a facility that spans eight football fields and holds over 17,700 picking locations. (cloud.google.com)) Across every shift, up to several dozen operators on ride‑on electric trucks crisscross the warehouse floor, with each operator navigating dozens of storage locations per tour. (cloud.google.com)) FM Logistic supplied AlphaEvolve with a seed program derived from their existing fast, cost‑prioritized allocation logic and a custom evaluation function built from a representative dataset of 60 tours (about one hour of workforce data), enabling the agent to test thousands of algorithm variants. (cloud.google.com)) AlphaEvolve pairs Gemini models in an evolutionary loop—using Gemini Flash for breadth and speed and Gemini Pro for depth—to generate mutated programs, run automated evaluators, and evolve the most promising solutions through recursive generations. (deepmind.google)) Google reports prior AlphaEvolve deployments recovered on average 0.7% of global compute capacity, sped up a critical Gemini training kernel by 23%, and reduced Gemini's training time by about 1%. (cloud.google.com)) The FM Logistic evaluation function enforced concrete operational constraints by penalizing exceeding forklift capacity, missing pending orders, assigning the same box twice, violating FIFO priority for older orders, and exceeding the compute time budget required for real‑time operations. (cloud.google.com)) FM Logistic says it is exploring extensions of the approach to other high‑volume e‑commerce facilities, to road transport routing for less‑than‑truckload shipments, and to AI‑driven product placement inside warehouses. (cloud.google.com))