Amazon Details 'DeepFleet' Robotics Framework
Amazon has published details on its DeepFleet framework, a multi-agent foundation model for coordinating warehouse robotics. The system focuses on enabling fleets of robots to dynamically allocate tasks and manage resources through agentic negotiation and autonomy. The architecture is designed to optimize for operational economics like throughput and labor cost, rather than just theoretical efficiency.
- The DeepFleet framework coordinates a fleet that surpassed 1 million mobile robots in June 2025, operating across more than 1,200 Amazon fulfillment and sortation centers globally. - This initiative builds on Amazon's 2012 acquisition of Kiva Systems for $775 million, which originally reduced "click to ship" cycle times from over 60 minutes to just 15. - DeepFleet is not a single entity but a suite of foundation models trained on billions of hours of operational data. The most promising architectures are a robot-centric (RC) transformer model and a graph-floor (GF) model that uses graph neural networks for spatial relationships. - The system uses generative AI to predict robot trajectories and forecast congestion, moving beyond reactive, rule-based pathfinding to predictive optimization. - Amazon projects DeepFleet will increase robot travel efficiency by 10%, contributing to faster order processing and lower costs. In new facilities equipped with the latest robotics, overall productivity has already increased by 25% compared to older sites. - The deployment comes as Amazon's robotic workforce approaches the size of its human one; the company employs around 1.56 million people and says it has retrained over 700,000 employees for more technical roles involving robotics. - The average number of human workers per fulfillment center has fallen to 670, its lowest point in 16 years, while the number of packages handled annually per worker has surged.
