DeepFleet Predicts Traffic Waves: Amazon’s AI for Coordinating a Million Robots
Amazon has crossed a major milestone: the company has deployed its one-millionth robot across fulfillment and sortation centers worldwide. At the same time, Amazon introduced DeepFleet, a suite of foundation models trained on billions of hours of real operational data to predict fleet behavior, minimize congestion, and improve robot efficiency by up to 10%.
Foundation models meet multi-robot systems
Foundation models transformed language and vision tasks by learning generalizable patterns from massive datasets. Amazon is applying the same principle to robotics, where forecasting interactions among thousands of mobile robots in dynamic warehouses requires more than rule-based planners or small-scale simulators. DeepFleet aims to generalize from diverse operational histories to forecast trajectories, anticipate congestion waves, and enable proactive control strategies.
Why prediction matters in warehouses
In fulfillment centers robots bring inventory shelves to human pickers; in sortation centers they move packages across complex floors. When fleets scale to tens or hundreds of thousands of agents, local jams, deadlocks, and cascading slowdowns become frequent. Predictive models let operators and controllers reroute, reschedule, or throttle traffic before problems cascade, reducing delays and improving throughput.
The four DeepFleet architectures
DeepFleet explores multiple inductive biases to model multi-robot dynamics at scale:
Robot-Centric (RC) Model: An autoregressive transformer focused on individual agents that uses local neighborhood observations such as nearby robots, objects, and floor markers. It handles asynchronous updates and integrates with a deterministic environment simulator to evolve states. With 97 million parameters, RC achieved the lowest errors in position and state predictions in evaluations.
Robot-Floor (RF) Model: This design combines robot states with global floor features like vertices and edges using cross-attention. It decodes actions synchronously and balances local interactions with warehouse-wide context. RF has about 840 million parameters and performed strongly on timing predictions.
Image-Floor (IF) Model: Treating the warehouse floor as a multi-channel image, IF uses convolutional encoders for spatial structure and transformers for temporal sequences. At scale it struggled to capture fine-grained, pixel-level robot interactions, which likely limited performance.
Graph-Floor (GF) Model: GF represents the environment as a spatiotemporal graph and blends graph neural networks with transformers. It models global relationships efficiently and, despite having just 13 million parameters, produced competitive predictions while remaining computationally lean.
These architectures vary in temporal handling (synchronous versus event-driven) and spatial perspective (local agent-centric versus global floor representations), letting Amazon test what scales best for large forecasting tasks.
Metrics, scaling, and compute tradeoffs
Evaluations used metrics such as dynamic time warping for trajectory accuracy and congestion delay error for operational realism. The RC model led overall with a position DTW of 8.68 and a congestion delay error of 0.11%. The GF model offered strong results at far lower parameter counts, highlighting tradeoffs between accuracy and computational cost.
Scaling studies mirrored patterns seen in other foundation models: larger models and richer datasets reduce prediction loss. For GF, projections suggest that a billion-parameter variant trained on millions of episodes could be an efficient use of compute. Amazon’s advantage is the sheer volume of real-world robot-hours available for training, which helps models generalize across layouts, robot generations, and operational cycles.
Early applications and operational impact
DeepFleet is already deployed across parts of Amazon’s global network of 300 plus facilities, including recent rollouts in Japan. Near-term uses include congestion forecasting, adaptive routing, and smarter task assignments. By improving travel efficiency and reducing bottlenecks, DeepFleet contributes to faster package processing and lower operational costs, with reported gains up to 10% in efficiency.
Amazon also emphasizes workforce development: since 2019 the company says it has upskilled more than 700,000 employees in robotics and AI-related roles, pairing automation with human jobs and shifting workers toward safer, higher-value tasks.
What comes next
Amazon plans to continue refining RC, RF, and GF variants, exploring scale and deployment tradeoffs. As foundation-model techniques extend from digital domains into physical automation, DeepFleet illustrates how predictive, data-driven control can move multi-robot systems from reactive coordination toward proactive, large-scale orchestration. This approach could reshape logistics and other industries that rely on dense fleets of mobile robots.