Research map / Architectures and efficiency / Attention and transformers
Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM
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.
Trajectory Mamba (Tamba) replaces attention in a motion-forecasting encoder-decoder with selective state-space models for linear-time complexity, adds joint polyline encoding and a Mamba-based cross decoder, and reports SOTA accuracy with fewer parameters and FLOPs on Argoverse 1 and 2.
Problem
Motion prediction for autonomous driving needs accurate multi-agent trajectory forecasts, but attention-based models scale quadratically with the number of targets and are costly for recursive, real-time prediction.
Approach
Redesign self-attention and cross-attention in a factorized encoder-decoder with selective SSM (Mamba) blocks for linear complexity, combine it with a joint polyline encoding that groups pedestrians and traffic lights in a shared embedder, and add an RNN-based weight inference module over K proposal trajectories.
Key results
- On Argoverse 2, Tamba reaches b-minFDE6 1.89 (vs. QCNet 1.91) with 4.54M parameters vs. QCNet's 7.66M and 27.3 GFLOPs vs. 45.3.
- On Argoverse 1, Tamba achieves b-minFDE6 1.67 vs. QCNet 1.69 with 4.54M parameters vs. 7.66M.
- Ablation on Argoverse 2 validation shows the Tamba variant reaches minFDE6 0.951 versus 1.094 for the attention variant (with joint encoding).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Tamba instead of QCNet | Trajectory forecasting · Argoverse 2 | minADE6: 0.65 -> 0.64 | Helped |
| Use Tamba instead of QCNet | Trajectory forecasting · Argoverse 2 | FLOPs (G): 45.3 -> 27.3 | Helped |
| Use Tamba instead of QCNet | Trajectory forecasting · Argoverse 1 | b-minFDE6: 1.69 -> 1.67 | Helped |
| Add joint polyline encoding to the attention variant (ablation) | Ours + Attention · Argoverse 2 validation set | minFDE6: 1.093 -> 1.094 | No clear effect |
| Add joint polyline encoding to the attention variant (ablation) | Ours + Attention · Argoverse 2 validation set | minADE6: 0.728 -> 0.763 | Hurt |
| Add joint polyline encoding to the Mamba variant (ablation) | Ours + Mamba · Argoverse 2 validation set | minADE6: 0.765 -> 0.724 | Helped |
| Add joint polyline encoding to the Tamba variant (ablation) | Ours + Tamba · Argoverse 2 validation set | minADE6: 0.680 -> 0.673 | Helped |
| Replace attention with Tamba (ablation) | Model with joint polyline encoding · Argoverse 2 validation set | minADE6: 0.763 -> 0.673 | Helped |
| Replace Mamba with Tamba (ablation) | Model with joint polyline encoding · Argoverse 2 validation set | minADE6: 0.724 -> 0.673 | Helped |
| Replace Mamba with Tamba (ablation) | Model without joint polyline encoding · Argoverse 2 validation set | minFDE6: 1.083 -> 0.955 | Mixed |
Limitations
- Some Argoverse 1 and 2 results could not be reproduced with full accuracy by the authors for certain baselines (marked '-').
- Ablation is limited to Argoverse 2 validation and no per-component efficiency numbers are given.
- Reported FLOPs reduction is inconsistent between abstract (four-fold) and body (about 40% faster in GigaOps).
Experimental setup
Datasets: Argoverse 1 Argoverse 2
Benchmarks: Argoverse 1 Argoverse 2
Metrics: minADE minFDE b-minFDE MR #Params FLOPs
Baselines: QCNet MTR GANet BANet HiVT Wayformer SceneTransformer LaneGCN DenseTNT THOMAS GoRela QML
Builds on: Mamba DETR QCNet HiVT
Closest papers by meaning
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.88)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.87)
- STraj: Self-training for Bridging the Cross-Geography Gap in Trajectory Prediction (AAAI 2025 · similarity 0.86)
- AutoTraces: Autoregressive Trajectory Forecasting via Multimodal Large Language Models (CVPR 2026 · similarity 0.86)
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.86)
- BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving (AAAI 2024 · similarity 0.85)
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.