CREST: Constraint-Release Execution for Multi-Robot Warehouse Shelf Rearrangement

A research team led by Jiaqi Tan has introduced CREST (Constraint-Release Execution for Multi-Robot Warehouse Shelf Rearrangement), a novel AI framework designed to optimize how teams of robots rearrange shelves in automated warehouses. The system tackles the Double-Deck Multi-Agent Pickup and Delivery (DD-MAPD) problem, a core challenge in logistics robotics. Unlike the previous state-of-the-art method, MAPF-DECOMP—which plans collision-free trajectories first and then rigidly assigns robots—CREST proactively releases trajectory constraints during execution. This key innovation allows robots to adapt on the fly, preventing them from idling while waiting for dependencies and reducing unnecessary trips to switch shelves.

Experiments across diverse warehouse layouts demonstrate CREST's significant performance gains. The framework consistently reduced key inefficiency metrics: agent travel distance by up to 40.5%, total completion time (makespan) by 33.3%, and costly shelf-switching actions by 44.4%. The benefits were even more pronounced when real-world factors like the time required to lift and place heavy shelves were modeled. These results highlight that execution-aware planning, where constraint management is dynamic, is crucial for scaling robotic warehouse operations. The team has made the code and data publicly available, providing a practical tool for improving the throughput and cost-effectiveness of automated fulfillment centers.