Research map / Label-efficient and robust learning / Recognition and long-tail
Bridging Past and Future: End-to-End Autonomous Driving with Historical Prediction and Planning
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 BridgeAD, an end-to-end autonomous driving framework that reformulates motion and planning queries as multi-step queries so historical prediction and planning can be injected step-wise into perception and motion planning, achieving SOTA on nuScenes open- and closed-loop evaluation.
Problem
End-to-end autonomous driving methods aggregate temporal information either via dense historical BEV features or sparse memory banks of trajectory queries, which omit historical information in motion planning or fail to match the multi-step nature of prediction and planning.
Approach
Represent motion and planning queries as multi-step queries (one per future time step), cache them in a FIFO memory queue, and use step-aligned cross-attention so that historical prediction supports current-frame perception while historical prediction and planning support future-frame motion planning, with a step-level Mot2Plan interaction for consistency.
Key results
- Open-loop planning on nuScenes: BridgeAD-B reaches 0.58 average L2 and 0.08% average collision rate, versus 0.61 L2 and 0.10% for SparseDrive.
- Closed-loop NeuroNCAP without post-processing: BridgeAD-S scores 1.52 with 76.2% collision rate versus SparseDrive 0.92 and 93.9%, and UniAD 0.73 and 88.6%.
- 3D detection on nuScenes val: BridgeAD-B achieves 0.507 mAP and 0.594 NDS, compared with 0.416 mAP for BEVFormer (R101).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add History-Enhanced Motion Prediction (ablation) | BridgeAD · nuScenes motion prediction | ADE (m), Car / Ped: 0.66 / 0.75 -> 0.62 / 0.70 | Helped |
| add Historical Mot2Det Fusion (ablation) | BridgeAD · nuScenes 3D detection | mAP: 0.404 -> 0.423 | Helped |
| add History-Enhanced Planning (ablation) | BridgeAD · nuScenes planning ablation | Avg. L2 (m): 0.71 -> 0.59 | Helped |
| add Step-Level Mot2Plan Interaction (ablation) | BridgeAD · nuScenes planning ablation | Avg. L2 (m): 0.68 -> 0.59 | Helped |
| use both step-level and mode-level self-attention instead of step-level only (ablation) | BridgeAD planning module · nuScenes planning ablation | Avg. L2 (m): 0.66 -> 0.59 | Helped |
| use both step-level and mode-level self-attention instead of mode-level only (ablation) | BridgeAD planning module · nuScenes planning ablation | Avg. L2 (m): 0.64 -> 0.59 | Helped |
| increase motion-history aggregation steps from 5 to 6 while holding planning-history steps at 3 (ablation) | BridgeAD · nuScenes planning ablation | Avg. L2 (m): 0.63 -> 0.59 | Helped |
| increase motion-history aggregation steps from 6 to 7 while holding planning-history steps at 3 (ablation) | BridgeAD · nuScenes planning ablation | Avg. L2 (m): 0.59 -> 0.62 | Hurt |
| increase planning-history aggregation steps from 2 to 3 while holding motion-history steps at 6 (ablation) | BridgeAD · nuScenes planning ablation | Avg. L2 (m): 0.64 -> 0.59 | Helped |
| increase planning-history aggregation steps from 3 to 4 while holding motion-history steps at 6 (ablation) | BridgeAD · nuScenes planning ablation | Avg. L2 (m): 0.59 -> 0.60 | Hurt |
Surprising findings
- Without self-attention in planning, historical information can only be aggregated for the planning steps, and removing step- or mode-level self-attention significantly degrades planning.
Limitations
- Historical prediction horizon for motion queries is shorter than the current-frame horizon, so only partial history is usable per step.
- Evaluated only on nuScenes and NeuroNCAP-based closed-loop simulation.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes NeuroNCAP
Metrics: L2 displacement error collision rate NeuroNCAP score mAP NDS AMOTA AMOTP ADE FDE MR
Baselines: UniAD VAD SparseDrive GenAD ST-P3 OccWorld Drive-WM ViP3D BEVFormer
Builds on: UniAD VAD SparseDrive MapTR
Closest papers by meaning
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.86)
- Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving (AAAI 2025 · similarity 0.84)
- AlignDrive: Aligned Lateral-Longitudinal Planning for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.84)
- Enhancing End-to-End Autonomous Driving with Latent World Model (ICLR 2025 · similarity 0.83)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.83)
- Continuously Learning, Adapting, and Improving: A Dual-Process Approach to Autonomous Driving (NeurIPS 2024 · similarity 0.83)
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.