Research map / Multi-agent systems and games / Human-AI and goal-conditioned
DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States
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.
Decouples multi-modal trajectory queries into mode queries (directional intentions) and state queries (time-evolving dynamic states), combining Attention and Mamba for motion forecasting on Argoverse 2 and nuScenes.
Problem
Motion forecasting for autonomous driving under the one-query-one-trajectory paradigm, which lacks detailed representation of how agent states evolve over future time steps, leading to suboptimal accuracy and temporal consistency.
Approach
Split each trajectory query into a mode query (directional intent, one per trajectory) and per-timestep state queries (dynamic states), process them with Attention and bidirectional Mamba modules, and fuse them in a Hybrid Coupling Module before decoding trajectories and probabilities.
Key results
- On the Argoverse 2 single-agent test set (single model), DeMo reports minFDE_1 3.74 and minADE_1 1.49 versus QCNet's 4.30 and 1.69, a 13.02% and 11.83% relative improvement.
- On the Argoverse 2 validation ablation, the full model reaches minFDE_6 1.268 and MR_6 0.152 versus 1.414 and 0.184 for the baseline without decoupled queries.
- DeMo runs in 38 ms per scene versus 94 ms for QCNet on an RTX 3090, with 5.9M parameters versus 7.7M for QCNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the one-query-one-trajectory approach with decoupled mode and state queries (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.414 -> 1.268 | Helped |
| use state queries only rather than the baseline mode-query setup (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.414 -> 1.505 | Hurt |
| add two auxiliary losses alongside decoupled queries (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.414 -> 1.405 | Helped |
| add the three aggregation modules while omitting auxiliary losses (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.414 -> 1.319 | Helped |
| use Bidirectional Mamba for state sequence modeling (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.307 -> 1.268 | Helped |
| replace Bidirectional Mamba with GRU for state sequence modeling (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.268 -> 1.842 | Hurt |
| replace Bidirectional Mamba with Unidirectional Mamba for state sequence modeling (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.268 -> 1.288 | Hurt |
| remove the state auxiliary loss (ablation) | DeMo · Argoverse 2 single-agent validation set | minADE6: 0.674 -> 0.715 | Hurt |
| remove the mode auxiliary loss (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.268 -> 1.289 | Hurt |
| remove the Hybrid Coupling Module (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.268 -> 1.324 | Hurt |
| remove the State Consistency Module (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.268 -> 1.317 | Hurt |
| remove the Mode Localization Module (ablation) | DeMo · Argoverse 2 single-agent validation set | minFDE6: 1.268 -> 1.297 | Hurt |
Surprising findings
- Directly adding state queries without the auxiliary losses and aggregation modules degrades performance (ID-2 is worse than the baseline), which the authors attribute to surplus queries whose meanings are hard to distinguish.
Limitations
- The decoupled query paradigm predicts longer trajectories with more queries, which can make the model heavier.
- The model is not designed with efficiency as a priority.
- Failure cases include missed turns into alleys and complex intersections, attributed to missing intent cues and imbalanced driving data.
Experimental setup
Datasets: Argoverse 2 nuScenes Argoverse 1 Waymo Open Motion Dataset
Benchmarks: Argoverse 2 nuScenes Argoverse 1 Waymo Open Motion Dataset
Metrics: minADE minFDE MR b-minFDE avgMinADE avgMinFDE actorMR
Baselines: QCNet SmartRefine MTR HPTR GANet ProphNet SIMPL THOMAS Forecast-MAE Gnet
Builds on: Mamba QCNet MTR VectorNet
Closest papers by meaning
- Motion Forecasting in Continuous Driving (NeurIPS 2024 · similarity 0.89)
- ModeSeq: Taming Sparse Multimodal Motion Prediction with Sequential Mode Modeling (CVPR 2025 · similarity 0.88)
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.88)
- UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.88)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.88)
- Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM (CVPR 2025 · similarity 0.88)
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.