AI Helps Warehouse Robots Work Smarter, Not Harder

In automated warehouses, teams of robots constantly rearrange inventory shelves to fulfill orders, but coordinating these movements efficiently has been a major . A new AI framework called CREST (Constraint-Release Execution of Shelf Trajectories) addresses this by improving how robots execute pre-planned shelf movements, leading to significant reductions in wasted travel time and unnecessary shelf handling. This advancement is crucial as warehouses scale up operations, where even small inefficiencies can compound into substantial delays and energy costs. The research, published in IEEE Robotics and Automation Letters, demonstrates that smarter execution strategies can boost throughput without requiring changes to warehouse infrastructure or robot hardware.

The key finding is that CREST consistently outperforms the previous state-of-the-art , MAPF-DECOMP, by proactively releasing constraints during execution. This allows robots to carry shelves more continuously, reducing idle time and minimizing how often shelves are switched between robots. In experiments across diverse warehouse layouts, CREST reduced metrics related to agent travel by up to 40.5%, makespan (total completion time) by up to 33.3%, and shelf switching by up to 44.4% compared to the baseline. These improvements were even greater when accounting for the time overhead of lifting and placing shelves, which is common in real-world operations.

Ology builds on a two-stage decomposition approach used in prior work. First, a Multi-Agent Path Finding (MAPF) solver computes collision-free trajectories for all shelves, ensuring they can move safely without interfering with each other. Second, CREST assigns robots to execute these trajectories but introduces three lightweight strategies to optimize the process: single trajectory replanning, dependency switching, and group trajectory replanning. These strategies leverage real-time execution information, such as when robots become available and when shelf constraints are cleared, to adjust plans dynamically. For example, dependency switching reverses precedence orders between shelves when beneficial, allowing a constrained shelf to proceed first without waiting.

From the paper show clear performance gains. On medium-sized layouts like R2R-M with 32 agents, CREST reduced the normalized sum of costs (a measure of non-carrying robot travel) from 7,575.40 to 4,727.12, a 37.6% improvement, and cut shelf switching from 1.48 to 0.82 per shelf. Figure 6 in the paper illustrates that applying all three strategies together achieved the best , with reductions of up to 40.5% in cost, 33.3% in makespan, and 44.4% in switching. The cost breakdown in Figure 7 reveals that most savings come from reducing agent travel without shelves, while shelf transport costs saw moderate decreases. Runtime analysis in Table III indicates planning is efficient, taking less than 0.06 seconds per shelf on medium layouts, making it feasible for real-time use.

Are significant for industries relying on automated warehouses, such as e-commerce and logistics, where faster rearrangement translates to quicker order fulfillment and lower operational costs. By minimizing robot idle time and shelf handling, CREST can enhance scalability in large facilities with hundreds of robots and thousands of shelves, as tested in layouts like S2W-L and DnE-L. The framework's robustness under lift/place overhead conditions further supports practical deployment, as it adapts to real-world timing constraints without sacrificing performance. This work underscores the value of execution-aware AI in robotics, moving beyond static planning to dynamic coordination that mirrors human-like flexibility.

Limitations noted in the paper include that CREST currently operates in an offline setting, though its low computational overhead suggests potential for online adaptation. The evaluation focused on well-formed instances where safe shelf plans exist, and while this covers many practical scenarios, extreme cases with highly congested layouts may require additional adjustments. Future work will explore extending CREST to dynamic environments with real-time task arrivals and validating it on physical multi-robot systems to confirm its effectiveness in live deployments.