Research map / Multi-agent systems and games / Human-AI and goal-conditioned
AgentFormer: Agent-Aware Transformers for Socio-Temporal Multi-Agent Forecasting
Ye Yuan, Xinshuo Weng, Yanglan Ou, Kris M. Kitani
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add agent-aware attention (vs standard scaled dot-product attention) (ablation) | AgentFormer · ETH/UCY average (K=20) | ADE_20/FDE_20: 0.27/0.47 -> 0.23/0.39 | Improved |
| add agent-aware attention (vs standard scaled dot-product attention) (ablation) | AgentFormer · nuScenes (K=5) | ADE_5/FDE_5: 2.02/4.29 -> 1.86/3.89 | Improved |
| use agent-aware attention instead of agent index encoding (ablation) | AgentFormer · ETH/UCY average (K=20) | ADE_20/FDE_20: 0.27/0.46 -> 0.23/0.39 | Improved |
| use agent-aware attention instead of agent index encoding (ablation) | AgentFormer · nuScenes (K=10) | ADE_10/FDE_10: 1.63/3.11 -> 1.45/2.86 | Improved |
| add joint latent intent modeling across agents (ablation) | AgentFormer · ETH/UCY average (K=20) | ADE_20/FDE_20: 0.27/0.46 -> 0.23/0.39 | Improved |
| add joint latent intent modeling across agents (ablation) | AgentFormer · nuScenes (K=5) | ADE_5/FDE_5: 1.95/3.98 -> 1.86/3.89 | Improved |
| add semantic map features (ablation) | AgentFormer · nuScenes (K=5) | ADE_5/FDE_5: 1.97/4.21 -> 1.86/3.89 | Improved |
| model time and social dimensions jointly instead of separate TF social + TF temporal (ablation) | AgentFormer · ETH/UCY average (K=20) | ADE_20/FDE_20: 0.26/0.47 -> 0.23/0.39 | Improved |
| model time and social dimensions jointly instead of separate TF social + TF temporal (ablation) | AgentFormer · nuScenes (K=5) | ADE_5/FDE_5: 1.99/4.12 -> 1.86/3.89 | Improved |
| use Transformer instead of LSTM as temporal model (GCN social) (ablation) | separate social/temporal baseline · ETH/UCY average (K=20) | ADE_20/FDE_20: 0.31/0.52 -> 0.28/0.50 | Improved |
| use Transformer instead of LSTM as temporal model (TF social) (ablation) | separate social/temporal baseline · nuScenes (K=10) | ADE_10/FDE_10: 1.69/3.31 -> 1.54/3.07 | Improved |
| add agent-aware attention (vs standard attention) (ablation) | AgentFormer · nuScenes (K=10) | FDE_10: 2.91 -> 2.86 | No clear effect |
Later papers that compared with this method
9 later accepted papers report AgentFormer as a baseline in their results tables; 7 of them report a better number on most of the results they share with it.
- Differentiable Semantic Meta-Learning Framework for Long-Tail Motion Forecasting in Autonomous Driving (AAAI 2026 · better on 1 of 1 shared result)
- SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction (CVPR 2025 · better on 1 of 2 shared results)
- SMEMO: Social Memory for Trajectory Forecasting (TPAMI 2024 · better on 3 of 3 shared results)
- VisionTrap: Vision-Augmented Trajectory Prediction Guided by Textual Descriptions (ECCV 2024 · better on 1 of 1 shared result)
- Joint Metrics Matter: A Better Standard for Trajectory Forecasting (ICCV 2023 · better on 5 of 7 shared results)
- IPCC-TP: Utilizing Incremental Pearson Correlation Coefficient for Joint Multi-Agent Trajectory Prediction (CVPR 2023 · better on 4 of 4 shared results)
- MUSE-VAE: Multi-Scale VAE for Environment-Aware Long Term Trajectory Prediction (CVPR 2022 · better on 5 of 7 shared results)
- Social ODE: Multi-agent Trajectory Forecasting with Neural Ordinary Differential Equations (ECCV 2022 · better on 3 of 6 shared results)
1 more not listed.
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SoPhie — SoPhie: An Attentive GAN for Predicting Paths Compliant to Social and Physical Constraints (CVPR 2019) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Scene Transformer: A unified architecture for predicting future trajectories of multiple agents (ICLR 2022 · similarity 0.92)
- Social ODE: Multi-agent Trajectory Forecasting with Neural Ordinary Differential Equations (ECCV 2022 · similarity 0.91)
- Latent Variable Sequential Set Transformers for Joint Multi-Agent Motion Prediction (ICLR 2022 · similarity 0.90)
- Poly-Autoregressive Prediction for Modeling Interactions (CVPR 2025 · similarity 0.87)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.87)
- BehaviorGPT: Smart Agent Simulation for Autonomous Driving with Next-Patch Prediction (NeurIPS 2024 · similarity 0.86)
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.
