Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Density-Adaptive Model Based on Motif Matrix for Multi-Agent Trajectory Prediction
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 DAMM, a density-adaptive trajectory predictor that uses a spatial-temporal motif matrix plus adaptive neighbor-range selection and interaction-probability weighting inside a CVAE-GAT framework, reporting gains on nuScenes and Argoverse.
Problem
Multi-agent trajectory prediction for autonomous driving ignores heterogeneous traffic density, which changes the number of interacting agents and their interaction probabilities over time.
Approach
Uses a higher-order motif matrix to encode agent interactions, adaptively selects multi-scale neighbor ranges in a spatial block (ASI), and estimates time-varying interaction probabilities with a pyramidal pooling temporal block (ATI); these form node and edge features for a GAT inside a CVAE.
Key results
- On nuScenes, ADE1 improves from 3.54 to 2.84 (reported as 19.77%), and FDE10 from 2.10 to 2.05 (2.38%).
- On Argoverse, maximum improvement of 9.52% in ADE (K=6) and 4.37% in FDE (K=1) over baselines.
- Ablation: adding spatial ASI and temporal ATI blocks to the motif-matrix baseline improves ADE1/FDE1 from 3.05/7.12 to 2.84/6.59.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAMM instead of the strongest listed baseline, Context-Aware | nuScenes trajectory prediction · nuScenes | ADE1: 3.54 -> 2.84 | Helped |
| Use DAMM instead of MUSE-VAE | nuScenes trajectory prediction · nuScenes | FDE10: 2.10 -> 2.05 | Helped |
| Use DAMM instead of R-Pred | Argoverse trajectory prediction · Argoverse | ADE1: 1.58 -> 1.57 | Helped |
| Use DAMM instead of ADAPT | Argoverse trajectory prediction · Argoverse | FDE6: 1.17 -> 1.29 | Hurt |
| Add the motif matrix, ASI, and ATI instead of using the adjacency-matrix interaction baseline (ablation) | DAMM on nuScenes · nuScenes | ADE1: 3.05 -> 2.84 | Helped |
| Add ASI, with motif matrix and ATI retained (ablation) | DAMM on nuScenes · nuScenes | ADE1: 2.97 -> 2.84 | Helped |
| Add ATI, with motif matrix and ASI retained (ablation) | DAMM on nuScenes · nuScenes | ADE1: 2.98 -> 2.84 | Helped |
| Use multi-scale neighbor selection instead of a single 50m radius (ablation) | DAMM on nuScenes · nuScenes | ADE1: 3.12 -> 2.84 | Helped |
| Add LSEL and LKLD to LDIS (ablation) | DAMM on nuScenes · nuScenes | ADE1: 2.99 -> 2.84 | Helped |
| Add LKLD to LDIS and LSEL (ablation) | DAMM on nuScenes · nuScenes | ADE1: 2.52 -> 2.84 | Hurt |
| Add LSEL to LDIS and LKLD (ablation) | DAMM on nuScenes · nuScenes | ADE1: 3.42 -> 2.84 | Helped |
Surprising findings
- The paper reports the best-of-K gains are largest at K=1 and shrink as K grows, which the authors attribute to increased prediction diversity.
Limitations
- Lane information is not further filtered or extracted, as the authors note.
- Performance is weaker than Context-Aware in the highest-density intervals (fourth and fifth), per the authors' density analysis.
- Percentage improvements are reported inconsistently between text and tables.
Experimental setup
Datasets: nuScenes Argoverse
Benchmarks: nuScenes Argoverse
Metrics: ADE FDE
Baselines: AgentFormer AutoBot MP++ THOMAS HiVT Aware GOHOME Context-Aware R-Pred ADAPT
Builds on: Graph Attention Network network motifs Spatial Pyramid Pooling CVAE
Closest papers by meaning
- TraPHic: Trajectory Prediction in Dense and Heterogeneous Traffic Using Weighted Interactions (CVPR 2019 · similarity 0.88)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.85)
- BehaviorGPT: Smart Agent Simulation for Autonomous Driving with Next-Patch Prediction (NeurIPS 2024 · similarity 0.85)
- Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling (NeurIPS 2025 · similarity 0.85)
- Reasoning Multi-Agent Behavioral Topology for Interactive Autonomous Driving (NeurIPS 2024 · similarity 0.85)
- BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.