Researchers from MIT and the technology firm Symbotic have developed a new artificial intelligence system designed to optimize traffic flow for fleets of warehouse robots. In busy autonomous warehouses, where hundreds of robots operate simultaneously, minor collisions or traffic jams can cause massive slowdowns. To prevent these inefficiencies, the team created a hybrid system that learns to prioritize robots dynamically, rerouting them in advance to avoid bottlenecks. The system combines deep reinforcement learning with a fast planning algorithm. Deep reinforcement learning allows the AI to learn through trial and error in simulations, determining which robots should move first based on developing congestion patterns. Once the AI decides on priorities, the planning algorithm calculates specific movement instructions for each robot. This combination enables the system to adapt rapidly to changing conditions, such as new orders or varying robot densities. In simulations inspired by actual e-commerce warehouse layouts, the new approach achieved a 25 percent increase in throughput compared to traditional methods. Throughput refers to the number of packages delivered per robot within a given time. The study emphasizes that even small percentage improvements in efficiency can have a significant financial impact on large-scale operations. Furthermore, the system demonstrated the ability to adapt to different warehouse layouts and robot quantities without requiring extensive reprogramming. Han Zheng, a graduate student at MIT's Laboratory for Information and Decision Systems and the lead author of the research, highlighted the limitations of current industry practices. Companies often rely on algorithms designed by human experts to manage robot movements. However, these static algorithms can fail when unexpected congestion occurs, sometimes forcing a complete shutdown of the warehouse to manually resolve issues. The new AI system addresses this by predicting future interactions and planning to avoid conflicts before they happen. The research paper, co-authored by Yining Ma, Brandon Araki, Jingkai Chen, and MIT professor Cathy Wu, was published in the Journal of Artificial Intelligence Research. Wu noted that the hybrid approach offers the best of both worlds by leveraging machine learning for complex decision-making while using classical optimization for precise path planning. She explained that pure machine learning often struggles with complex optimization, while human-designed algorithms are labor-intensive and inflexible. The integration of both methods simplifies the learning task for the AI. Although the system is not yet ready for real-world deployment, the results demonstrate the feasibility of using machine learning to guide warehouse automation. The researchers plan to expand the system in the future to include task assignments, which currently are determined separately from movement planning. They also intend to scale the technology to handle larger warehouses with thousands of robots. The project was funded by Symbotic.