DeepFleet: Amazon's AI Models Predict Robot Traffic, Boost Efficiency
Amazon has achieved a significant milestone, deploying its one-millionth robot across global fulfillment and sortation centers, solidifying its position as the world’s largest operator of industrial mobile robotics. This remarkable expansion coincides with the launch of DeepFleet, a pioneering suite of AI foundation models engineered to enhance coordination among these vast robot fleets. Trained on billions of hours of real-world operational data, these models are poised to optimize robot movements, significantly reduce congestion, and boost overall efficiency by up to 10%.
The concept of foundation models, which gained prominence in language and vision AI, involves training massive datasets to learn general patterns that can then be adapted to a multitude of specific tasks. Amazon is now applying this powerful approach to robotics, where the challenge of coordinating thousands of robots in dynamic warehouse environments demands a level of predictive intelligence far beyond what traditional simulations can offer. In fulfillment centers, robots are essential for transporting inventory shelves to human workers, while in sortation facilities, they efficiently handle packages destined for delivery. With fleets numbering in the hundreds of thousands, operational bottlenecks like traffic jams and deadlocks are common, slowing down the entire process. DeepFleet directly addresses these issues by accurately forecasting robot trajectories and interactions, enabling proactive planning and intervention.
The models leverage an incredibly rich and diverse dataset, spanning millions of robot-hours and encompassing various warehouse layouts, robot generations, and operational cycles. This extensive data allows DeepFleet to capture complex emergent behaviors, such as congestion waves, and generalize across diverse scenarios, much like how large language models adapt to new queries.
DeepFleet is built upon four distinct model architectures, each designed with a unique approach to understanding multi-robot dynamics. The Robot-Centric (RC) model, for instance, functions like a focused observer, predicting individual robot actions based on local neighborhood data, such as nearby robots, objects, and markers. Despite its relatively modest size with 97 million parameters, this model demonstrated superior accuracy in position and state predictions during evaluations. In contrast, the Robot-Floor (RF) model takes a broader view, integrating individual robot states with global floor features like vertices and edges, enabling synchronous predictions that balance local interactions with warehouse-wide context. This larger model, with 840 million parameters, performed strongly on timing predictions. A third approach, the Image-Floor (IF) model, attempted to visualize the warehouse as a multi-channel image using convolutional encoding for spatial features, but it underperformed, likely due to difficulties in capturing precise, pixel-level robot interactions at scale. Finally, the Graph-Floor (GF) model offers a computationally efficient solution, representing the warehouse floor as a spatiotemporal graph. This allows it to handle global relationships efficiently, predicting actions and states with only 13 million parameters, making it lean yet highly competitive. These varied designs, differing in their temporal (synchronous versus event-based) and spatial (local versus global) approaches, allow Amazon to test which methods are best suited for large-scale forecasting.
Performance evaluations on unseen warehouse data used metrics such as dynamic time warping (DTW) for trajectory accuracy and congestion delay error (CDE) for operational realism. The RC model led overall, achieving a DTW score of 8.68 for position and a CDE of 0.11%, while the GF model offered strong results with significantly lower computational complexity. Scaling experiments further confirmed that larger models trained on more extensive datasets consistently reduce prediction losses, mirroring trends observed in other foundation models. For the GF model, extrapolations suggest that a 1-billion-parameter version, trained on 6.6 million episodes, could achieve optimal computational efficiency. This scalability is a critical advantage, as Amazon’s vast robot fleet provides an unparalleled volume of operational data. Early applications of DeepFleet include congestion forecasting and adaptive routing, with future potential extending to automated task assignment and deadlock prevention.
DeepFleet is already making a tangible impact on Amazon’s global network, which spans over 300 facilities worldwide, including recent deployments in Japan. By improving robot travel efficiency, the technology directly contributes to faster package processing and reduced operational costs, ultimately benefiting customers. Beyond efficiency, Amazon also highlights its commitment to workforce development, having upskilled over 700,000 employees since 2019 in robotics and AI-related roles. This integration aims to create safer jobs by offloading physically demanding tasks to machines.
As Amazon continues to refine DeepFleet, focusing on its most promising RC, RF, and GF variants, this technology could redefine multi-robot systems in logistics and beyond. By leveraging advanced AI to anticipate fleet behaviors, DeepFleet moves beyond reactive control, paving the way for more autonomous and scalable operations. This innovation underscores how foundation models are extending their influence from purely digital realms into physical automation, poised to transform industries reliant on coordinated robotics.