Research map / Graph learning / Knowledge and temporal graphs
Learning Cooperative Trajectory Representations for Motion Forecasting
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 V2X-Graph, a graph-based model that fuses cooperative trajectories (motion and interaction features) end-to-end for cooperative motion forecasting, and releases the V2X-Traj V2V&I forecasting dataset.
Problem
Cooperative motion forecasting for autonomous driving using multi-view historical trajectories from infrastructure and other vehicles; existing cooperative methods only fuse perception at single frames and underuse motion and interaction context.
Approach
Represent cooperative trajectories as graph nodes with motion and spatial-temporal encodings, then use an interpretable association (link prediction supervised by IoU-based pseudo labels) to guide three subgraphs (motion fusion, agent-lane, cooperative interaction) that fuse features across views end-to-end.
Key results
- On V2X-Seq, V2X-Graph with feature fusion reaches minADE 1.05, minFDE 1.79, MR 0.25, beating PP-VIC+HiVT (1.12/1.97/0.30).
- On V2X-Traj V2V&I setting, V2X-Graph achieves minADE 0.72, minFDE 1.13, MR 0.11, outperforming HDGT (0.94/1.56/0.17).
- Ablation on V2X-Traj validation shows removing any of MFG, ALG, or CIG degrades minFDE (e.g., 1.79 full vs 2.60 without MFG).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fuse cooperative trajectory features instead of using vehicle-only trajectories | V2X-Graph · V2X-Seq | minADE: 1.16 -> 1.05 | Helped |
| Add MFG to the full method (ablation) | V2X-Graph · V2X-Traj validation | minADE: 1.16 -> 0.95 | Helped |
| Add ALG to the full method (ablation) | V2X-Graph · V2X-Traj validation | minADE: 1.26 -> 0.95 | Helped |
| Add CIG to the full method (ablation) | V2X-Graph · V2X-Traj validation | minADE: 1.08 -> 0.95 | Helped |
| Add MFG to cooperative feature representations (ablation) | V2X-Graph · V2X-Traj validation | minADE: 1.03 -> 0.95 | Helped |
| Add ALG to cooperative feature representations (ablation) | V2X-Graph · V2X-Traj validation | minADE: 1.19 -> 0.95 | Helped |
| Add CIG to cooperative feature representations (ablation) | V2X-Graph · V2X-Traj validation | minADE: 1.08 -> 0.95 | Helped |
| Use interpretable rather than full feature fusion (ablation) | V2X-Graph · V2X-Traj validation | minFDE: 2.09 -> 1.79 | Helped |
| Increase the proportion of disturbed pseudo labels (ablation) | V2X-Graph · V2X-Traj validation | motion forecasting performance | Hurt |
| Drop more infrastructure-view perception data (ablation) | V2X-Graph · V2X-Seq | minADE: 1.05 -> 1.10 | Hurt |
Surprising findings
- Perception-completion (PP-VIC) provided to DenseTNT degrades its performance on V2X-Seq, indicating error propagation from cross-view perception.
Limitations
- Performance depends on tracking quality from each view.
- Joint perception-to-forecasting optimization left to future work.
- Cooperative data collection is costly, so large-scale cooperative training data is limited (scalability study on V2X-Seq subsets).
Experimental setup
Datasets: V2X-Seq V2X-Traj
Benchmarks: V2X-Seq V2X-Traj
Metrics: minADE minFDE MR
Baselines: PP-VIC + DenseTNT PP-VIC + HiVT DenseTNT HiVT HDGT
Builds on: HiVT DenseTNT HDGT PP-VIC
Closest papers by meaning
- SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction (ICLR 2025 · similarity 0.87)
- End-to-End Autonomous Driving Through V2X Cooperation (AAAI 2025 · similarity 0.86)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.86)
- Motion Forecasting in Continuous Driving (NeurIPS 2024 · similarity 0.86)
- UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.86)
- SEPT: Towards Efficient Scene Representation Learning for Motion Prediction (ICLR 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.