Amazon Robotics' DeepFleet Predicts How Thousands of Warehouse Robots Move
Amazon Robotics researchers publish a paper comparing four foundation model architectures trained on warehouse robot movement data, finding local, event-based designs scale best.
ProTakeaways
- Amazon Robotics researchers publish DeepFleet, a suite of four foundation model architectures for warehouse robot fleet forecasting.
- Compared architectures: robot-centric transformer, robot-floor cross-attention, image-floor CNN, and graph-floor GNN with edge-conditioned attention.
- Robot-centric (97M params) achieved best overall accuracy; graph-floor (13M) was competitive at 60x smaller size.
- Image-based whole-floor representation performed worst, suggesting pixel-level convolutions poorly capture robot interactions.
- Scaling experiments derived power-law fit for graph-floor: 1B model needs ~6.6M floor episodes at 10^22 FLOPs.
- Trained on production data from hundreds of thousands of Amazon warehouse robots; models and code not released publicly.
A team at Amazon Robotics has released a research paper introducing DeepFleet, a suite of four foundation model architectures for forecasting how large fleets of warehouse robots move and interact. The work is a systematic study of design choices rather than a product launch: the researchers train each architecture on production movement data from Amazon fulfillment and sortation centers, then compare how their inductive biases hold up on trajectory prediction and congestion metrics.
The core question the paper tries to answer is which representational bets pay off when you scale multi-agent forecasting to thousands of robots per floor. Each of the four architectures embodies a distinct inductive bias and collectively they explore key points in the design space for multi-agent foundation models. The robot-centric (RC) model is an autoregressive decision transformer operating on neighborhoods of individual robots. The robot-floor (RF) model uses a transformer with cross-attention between robots and the warehouse floor. The image-floor (IF) model applies convolutional encoding to a multi-channel image representation of the full fleet. The graph-floor (GF) model combines temporal attention with graph neural networks for spatial relationships.
Why warehouse fleets break traditional planners
Historically, warehouse coordination has been the domain of Multi-Agent Path Finding (MAPF), an area dominated by heuristic search algorithms like conflict-based search and priority inheritance with backtracking. These methods work well on abstract robot models, but the Amazon team argues they struggle once you fold in real dynamics, physical constraints, and hundreds of agents whose interactions produce emergent phenomena like traffic waves and deadlocks.
DeepFleet borrows a different bet from the language and vision playbook: pretrain on massive real-world data with a self-supervised objective (next-state and next-action forecasting), and hope the resulting representations transfer to downstream tasks like congestion forecasting, adaptive routing, and rescheduling. The paper frames forecasting as the pretraining objective, with control as future work.