Skip to content

Heteroscedastic Diffusion for Multi-Agent Trajectory Modeling

Home / Papers / Heteroscedastic Diffusion for Multi-Agent Trajectory Modeling

Heteroscedastic Diffusion for Multi-Agent Trajectory Modeling

Authors: Guillem Capellera, Antonio Rubio, Luis Ferraz, Antonio Agudo Date: 2026-05-11 Paper ID: openalex:2605.10717

Summary

This paper introduces U2Diffine, a diffusion-based framework designed to solve both multi-agent trajectory forecasting and completion tasks. The method augments standard denoising objectives with a negative log-likelihood loss for noise, allowing for the propagation of uncertainty to real state space via first-order Taylor approximations. Additionally, the authors provide U2Diff for accelerated sampling and a Rank Neural Network to rank multi-modal outputs by estimated error probability. Experimental results demonstrate state-of-the-art performance across multiple sports trajectory benchmarks.

Key Contributions

  • Introduced U2Diffine, a unified diffusion model for trajectory completion and forecasting that provides state-wise heteroscedastic uncertainty.
  • Developed a noise-based NLL loss with first-order Taylor propagation to map latent space uncertainty to the state space.
  • Proposed U2Diff for efficient, gradient-free sampling and a RankNN module to estimate error probabilities for multi-modal trajectories.
  • Achieved state-of-the-art performance on four sports trajectory datasets (NBA, Basketball-U, Football-U, Soccer-U).

Open Questions & Future Work

Key Concepts

  • u2diffine: A diffusion-based framework for multi-agent trajectory completion and forecasting that provides state-wise heteroscedastic uncertainty via latent space uncertainty propagation.

Archivist Review

I approved U2Diffine as a distinct mechanism for combining diffusion with state-wise uncertainty quantification, and a consolidated open question regarding the need for flexible temporal/spatial trajectory modeling. I rejected the user’s initial open question slug because it was redundant with the one I approved. Sports datasets were excluded as they are routine in this domain and lack the uniqueness required for standalone vault entries.

Approved Concepts

  • U2Diffine: It enables unified handling of trajectory completion and forecasting with explicit heteroscedastic uncertainty quantification through diffusion processes.

Approved Open Questions

  • Flexible Multi-Agent Trajectory Modeling: Flexibility in temporal and spatial dimensions is critical for applying trajectory modeling to varying, real-world sports and surveillance scenarios where observation durations and agent counts change dynamically.

Rejected Candidates

  • [open_question] Flexible Multi-Agent Trajectory Modeling (variable-temporal-spatial-trajectory-modeling) - duplicate_existing: This is a near-duplicate of the approved open question; the approved version uses a more descriptive and professional slug.

Metadata & Links

url
https://arxiv.org/abs/2605.10717
paper_id
2605.10717
paper_source
openalex
domain
time-series
tags
diffusion-modelstrajectory-forecastinguncertainty-estimationmulti-agent-systems
architectures
datasets
concept_slugs
u2diffine
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-14T06:10:38Z