Research map / 3D vision / 3D generation
DualAD: Disentangling the Dynamic and Static World for End-to-End Driving
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 DualAD, a dual-stream transformer that represents dynamic agents with object queries and static scene with BEV grid queries, linked by dynamic-static cross-attention, and integrates it into UniAD and VAD for end-to-end driving on nuScenes.
Problem
Unified BEV-grid belief states cannot explicitly compensate for object motion and handle highly dynamic agents well, which hurts perception, tracking, and downstream prediction and planning in end-to-end autonomous driving.
Approach
Disentangle the scene into an object-centric dynamic stream (object queries that attend directly to image features and are propagated with ego and object motion compensation) and a BEV-grid static stream (ego-motion compensation only), coupled by a novel dynamic-static cross-attention where object queries attend to nearby BEV queries.
Key results
- DUALAD-I reaches 49.5 mAP and 57.8 NDS on nuScenes val object detection, +1.3 mAP over StreamPETR (48.2).
- DUALAD-I achieves 55.1 AMOTA, +15.8 over UniAD-I (39.3) and +2.5 over StreamPETR (52.6).
- Integrated into VAD, DUALVAD improves vehicle motion prediction EPA from 64.7 to 69.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace UniAD with the DUALAD dual-stream design | UniAD · nuScenes val object detection | mAP: 39.5 -> 49.5 | Helped |
| replace StreamPETR with DUALAD-I | StreamPETR · nuScenes val object detection | mAP: 48.2 -> 49.5 | Helped |
| replace UniAD-I with DUALAD-I | UniAD-I · nuScenes val map segmentation | lane IoU (%): 31.3 -> 34.6 | Helped |
| replace StreamPETR with DUALAD-I | StreamPETR · nuScenes val tracking | AMOTA: 52.6 -> 55.1 | Helped |
| replace UniAD with DUALAD | UniAD · nuScenes val, vehicle motion prediction | EPA: 45.6 -> 52.4 | Helped |
| replace UniAD with DUALAD | UniAD · nuScenes val, pedestrian motion prediction | EPA: 35.5 -> 45.2 | Helped |
| replace UniAD with DUALAD | UniAD · nuScenes val, 3s ego planning | collision rate (%): 0.71 -> 0.35 | Helped |
| add dynamic-static cross-attention (ablation) | DUALAD-I · nuScenes val object detection | mAP: 47.7 -> 49.5 | Helped |
| add dynamic-static cross-attention (ablation) | DUALAD-I · nuScenes val tracking | AMOTA: 51.9 -> 55.1 | Helped |
| make stream interaction bidirectional (ablation) | DUALAD-I · nuScenes val object detection | mAP: 49.5 -> 49.3 | No clear effect |
| remove temporal attention for BEV queries (ablation) | DUALAD-I · nuScenes val object detection | mAP: 49.5 -> 46.9 | Hurt |
| remove temporal attention for the unified BEV grid (ablation) | UniAD-I† · nuScenes val object detection | NDS: 51.2 -> 54.9 | Helped |
Surprising findings
- Removing temporal attention from the unified BEV grid in UniAD improves its dynamic agent perception (+3.7 NDS, +4.4 AMOTA), contrary to the expected benefit of temporal modeling.
- Bidirectional stream interaction (static queries attending to dynamic agents) does not yield significant improvements over one-directional interaction.
Limitations
- Perception performance slightly degrades when training in the end-to-end setting.
- Open-loop planning results are of questionable validity on nuScenes, as the authors note.
- Only camera input is used; LiDAR integration is left as future work.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: mAP NDS mATE mAOE mAVE IoU AMOTA AMOTP IDS EPA
Baselines: StreamPETR BEVFormer UniAD VAD PF-Track ImmortalTrack+ BEVerse VIP3D
Builds on: StreamPETR UniAD VAD BEVFormer
Closest papers by meaning
- Continuously Learning, Adapting, and Improving: A Dual-Process Approach to Autonomous Driving (NeurIPS 2024 · similarity 0.88)
- DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Driving (ICLR 2025 · similarity 0.88)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.87)
- Navigation-Guided Sparse Scene Representation for End-to-End Autonomous Driving (ICLR 2025 · similarity 0.87)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.87)
- VR-Drive: Viewpoint-Robust End-to-End Driving with Feed-Forward 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.87)
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.