Research map / Multi-agent systems and games / Human-AI and goal-conditioned
HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention
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 HPNet, a dynamic trajectory forecaster that attends over historical prediction embeddings (Historical Prediction Attention) alongside agent and mode attention, improving accuracy and temporal stability on Argoverse and INTERACTION.
Problem
Trajectory forecasting for autonomous driving treats each prediction independently from a fixed-length history window, causing unstable and temporally inconsistent successive predictions and a limited visible temporal range.
Approach
Treats trajectory prediction as a dynamic task by letting current prediction embeddings attend to historical prediction embeddings (not trajectories, to keep training parallel), which both stabilizes successive predictions and extends the effective observation window without extra computation; combined with agent and mode attention as a Triple Factorized Attention module, and a refinement pass fee
Key results
- On Argoverse test, HPNet single model achieves b-minFDE 1.7375, minFDE 1.0986, MR 0.1067, minADE 0.7612, best among single models.
- On INTERACTION test, HPNet single model reaches minJointFDE 0.8231 vs. FJMP 0.9218 and minJointADE 0.2548 vs. 0.2752.
- Ablation on Argoverse validation: removing Historical Prediction Attention raises b-minFDE from 1.506 to 1.527; removing Triple Factorized Attention raises it by 0.326.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the full Triple Factorized Attention module (ablation) | HPNet · Argoverse validation | b-minFDE: 1.832 -> 1.506 | Helped |
| Add Agent Attention (ablation) | HPNet · Argoverse validation | b-minFDE: 1.711 -> 1.506 | Helped |
| Add Historical Prediction Attention (ablation) | HPNet · Argoverse validation | b-minFDE: 1.527 -> 1.506 | Helped |
| Add Mode Attention (ablation) | HPNet · Argoverse validation | b-minFDE: 1.531 -> 1.506 | Helped |
| Add Historical Prediction Attention (ablation) | HPNet · Argoverse validation, predictions across time steps 20–30 | summed ADE of overlapping successive predictions: about 2.90 -> about 2.25 | Helped |
| Add Historical Prediction Attention (ablation) | HPNet · Argoverse validation, predictions across time steps 20–30 | b-minFDE | Helped |
| Add Historical Prediction Attention (ablation) | HPNet · Argoverse validation, abrupt right-turn scenario | reaction timeliness | No clear effect |
Surprising findings
- Ensembling yields only a small minFDE gain for HPNet (0.013), which the authors attribute to its more stable predictions.
Limitations
- Ensemble gains are small, suggesting limited benefit from ensembling.
- Relies on prediction embeddings rather than trajectories for history, a design choice made for training parallelism.
Experimental setup
Datasets: Argoverse INTERACTION
Benchmarks: Argoverse INTERACTION
Metrics: minADE minFDE MR b-minFDE minJointADE minJointFDE
Baselines: LaneGCN mmTransformer DenseTNT THOMAS TPCN SceneTransformer HiVT GANet Multipath++ Macformer DCMS HeteroGCN
Builds on: QCNet DCMS LaneGCN DETR
Closest papers by meaning
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.88)
- Motion Forecasting in Continuous Driving (NeurIPS 2024 · similarity 0.87)
- Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM (CVPR 2025 · similarity 0.87)
- STraj: Self-training for Bridging the Cross-Geography Gap in Trajectory Prediction (AAAI 2025 · similarity 0.87)
- SmartRefine: A Scenario-Adaptive Refinement Framework for Efficient Motion Prediction (CVPR 2024 · similarity 0.86)
- LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations (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.