Research map / Multi-agent systems and games / Human-AI and goal-conditioned

Unified Uncertainty-Aware Diffusion for Multi-Agent Trajectory Modeling

CVPR 2025 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Interactive multi-agent trajectory prediction · 10 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes U2Diff, a unified diffusion model for multi-agent trajectory completion that jointly predicts state-wise uncertainty via an NLL loss augmentation and ranks sampled modes with a supervised RankNN to estimate per-mode error probabilities.

Problem

Multi-agent trajectory modeling has focused on forecasting, lacking unified trajectory completion, state-wise uncertainty estimates, and per-mode error probabilities for ranking multi-modal samples under the same prior.

Approach

Augment the DDPM denoising loss with the Gaussian negative log-likelihood of the predicted noise to learn per-state variance, propagate latent variance to state space during DDIM sampling, and add a Rank Neural Network (RankNN) that outputs softmax error probabilities over K sampled scene modes, trained with a differentiable Spearman correlation objective.

Key results

• On trajectory completion, U2Diff reduces minADE20 on Football-U from 3.55 (UniTraj) to 2.42 and on Soccer-U from 94.59 to 53.93, and gives about 27% better minSADE20 on Basketball-U.
• On NBA forecasting, U2Diff achieves minSADE20 1.48 vs 1.63 for LED, over 9% better in scene-level metric.
• RankNN achieves median Spearman correlation with SADE of 0.58 (NBA), 0.78 (Soccer-U) range, versus 0.25-0.35 for AvgUcty on the NBA and other datasets.

Measured effects reported in the paper

Surprising findings

• Omitting the NLL loss augmentation (lambda=0) does not degrade minSADE20 performance, though it helps minADE20 under the UniTraj metric.
• Variance propagation works best when started at an intermediate denoising step (30 of 50) rather than from the first step.

Limitations

• NBA long-horizon forecasting shows lower AccRate and higher NLL, indicating weaker uncertainty calibration over long horizons.
• Basketball-U with five or more fully unseen agents is more challenging, with slightly lower AccRate.
• Covariance across state dimensions is approximated as null during variance propagation to avoid cost and instability.

Experimental setup

Datasets: NBA SportVU Basketball-U Football-U Soccer-U

Benchmarks: Basketball-U Football-U Soccer-U NBA

Metrics: minADE20 minFDE20 minSADE20 minSFDE20 AccRate NLL Spearman correlation minSADEk

Baselines: UniTraj LED AutoBots GroupNet MID SSSD GC-VRNN Naomi

Builds on: DDPM DDIM CSDI Mamba UniTraj

Closest papers by meaning

• Sports-Traj: A Unified Trajectory Generation Model for Multi-Agent Movement in Sports (ICLR 2025 · similarity 0.87)
• Improving Transferability for Cross-Domain Trajectory Prediction via Neural Stochastic Differential Equation (AAAI 2024 · similarity 0.86)
• Proper Scoring Rules for Agentic Uncertainty Quantification (NeurIPS 2026 · similarity 0.86)
• Event2Tracking: Reconstructing Multi-Agent Soccer Trajectories Using Long-Term Multimodal Context (AAAI 2025 · similarity 0.86)
• C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction (AAAI 2025 · similarity 0.86)
• SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.86)