Reproduce “FlowMaps: Modeling Long-Term Multimodal Object Dynamics with Flow Matching” 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

Francesco Argenziano, Miguel Saavedra-Ruiz, Sacha Morin, Charlie Gauthier, Daniele Nardi, Liam Paull

FlowMaps models long-term multimodal object dynamics by leveraging a latent flow matching approach to predict continuous 3D object placements in dynamic environments. The model achieves superior performance in object navigation tasks across diverse human habits, demonstrating strong generalization to unseen environments and validated through real-world robot deployment.

Key insights

  • Flow matching effectively learns continuous, multimodal spatio-temporal distributions of object placements, capturing latent regularities induced by human routines without requiring explicit activity identification.
  • A hierarchical architecture combining a Map Encoder for scene context aggregation and CDiT blocks for query-specific flow refinement efficiently models complex scene dependencies and object dynamics.
  • The approach demonstrates strong generalization to unseen environments and distinct human habits, indicating that learned object motion patterns are transferable beyond specific training households.

Summary by alphaXiv.

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