Reproduce “DeepFleet: Multi-Agent Foundation Models for Mobile Robots” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ameya Agaskar, Sriram Siva, William Pickering, Kyle O'Brien, Charles Kekeh, Alexandre Ormiga Galvao Barbosa, Ang Li, Brianna Gallo Sarker and 13 more

Amazon Robotics developed DEEPFLEET, a suite of multi-agent foundation models designed for large-scale mobile robot fleets, trained on extensive real-world operational data. The Robot-Centric (RC) architecture demonstrated superior performance in predicting future robot positions and states, achieving a Dynamic Time Warping (DTW) position of 8.68 and a Congestion Delay Error (CDE) of 3.40% over a 60-second horizon.

Key insights

  • Models that predict discrete robot actions (e.g., Robot-Centric), which then evolve the environment state, generally outperform models directly predicting global floor states (e.g., Image-Floor).
  • Image-based global representations, when applied without a strong inductive bias for granular robot interactions, are less effective for modeling complex fleet dynamics.
  • Architectures focusing on local interaction propagation (Robot-Centric) or explicit spatiotemporal relational graphs (Graph-Floor) achieved superior performance and parameter efficiency compared to models providing full spatial context to every robot.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.