New SRA algorithm optimizes warehouse layouts 80x faster without simulations

Automated warehouses coordinate hundreds to thousands of robots this paper tackles the problem of optimizing physical layouts the placement of storage shelves to maximize throughput. Previous state-of-the-art approaches rely on evolutionary optimization methods, which treat the warehouse as a black box and use random mutation to iteratively improve layouts. Though effective, these methods are sample-inefficient—they require massive numbers of simulations to evaluate each candidate layout, making them impractical for rapid reconfiguration.

Stress-Relief Annealing (SRA), introduced by researchers from Kyushu University and Carnegie Mellon, replaces this brute-force approach with a mathematically grounded algorithm. It converts task demand into a per-vertex stress field that predicts where traffic will concentrate, with the field's peak provably capping throughput. SRA then anneals the layout to relieve those stress peaks. In experiments, SRA roughly doubled the number of robots a human-designed warehouse could sustain, matched or exceeded evolutionary baseline throughput, and did so in only 19 minutes on a single CPU core—compared to 25,000 simulations and 25 hours on a 64-core machine. The gains generalized across different Multi-Agent Path Finding algorithms, non-uniform task demands, and a twice-sized warehouse.