Research map / Label-efficient and robust learning / Recognition and long-tail
Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous 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 MomAD, an end-to-end driving planner that adds trajectory momentum (Hausdorff-distance trajectory matching) and perception momentum (LSTM-based historical query mixing) plus instance denoising, evaluated on nuScenes, a curated Turning-nuScenes, and Bench2Drive.
Problem
End-to-end driving planners predict one-shot multi-modal trajectories from single-frame perception, causing temporal inconsistency (unstable control, trembling) and vulnerability to occlusions.
Approach
Introduce momentum into end-to-end planning: select the candidate trajectory closest (Hausdorff distance) to the previous plan, then cross-attend that planning query with historical planning queries (LSTM-processed, score-weighted) to refine the trajectory; add instance-feature noise denoising during training and a TPC consistency metric.
Key results
- On Turning-nuScenes over a 6s horizon, MomAD reduces collision rate by 26% and TPC by 0.97m (33.45%) versus SparseDrive.
- On Bench2Drive closed-loop, MomAD reaches 51.20 driving score vs 48.63 for SparseDrive and improves success rate by up to 16.3% over VAD-mmt.
- On nuScenes planning, MomAD achieves avg L2 0.60m, collision rate 0.09%, and TPC 0.54m, versus SparseDrive's 0.61m, 0.08%, and 0.57m.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MomAD momentum planning to SparseDrive | SparseDrive · nuScenes val | TPC (m) Avg.: 0.57 -> 0.54 | Helped |
| add MomAD momentum planning to SparseDrive | SparseDrive · Turning-nuScenes val | Collision rate (%) Avg.: 0.40 -> 0.32 | Helped |
| add MomAD momentum planning to SparseDrive | SparseDrive · Turning-nuScenes val, 6s horizon | TPC (m): 2.90 -> 1.93 | Helped |
| add MomAD to VAD | VAD · Bench2Drive closed-loop | Success rate (%): 15.00 -> 17.44 | Helped |
| add Encoder and Denoise (ED) (ablation) | MomAD sparse perception · nuScenes val | mAP: 0.407 -> 0.420 | Helped |
| set ED Gaussian noise factor NS to 0.1 (ablation) | MomAD sparse perception · nuScenes val | L2 (m) Avg.: 0.62 -> 0.60 | Helped |
| set ED Gaussian noise factor NS to 0.2 (ablation) | MomAD sparse perception · nuScenes val | Collision rate (%) Avg.: 0.09 -> 0.18 | Hurt |
| use two frames of history instead of empty history (ablation) | MomAD momentum planning · Turning-nuScenes val | TPC (m) Avg.: 0.81 -> 0.63 | Helped |
| use three frames of history instead of two (ablation) | MomAD momentum planning · Turning-nuScenes val | L2 (m) Avg.: 0.76 -> 0.78 | Hurt |
| replace Long-horizon Query Mixer with Add (ablation) | MomAD momentum planning · Turning-nuScenes val | TPC (m) Avg.: 0.63 -> 0.67 | Hurt |
Surprising findings
- Adding the momentum (MP) module at t=1 with empty history does not improve performance; gains appear only when historical results are used.
- Using more history frames (t=3) reduces the improvement relative to t=2, attributed to uncertainty from departing historical features.
Limitations
- Mode collapse from teacher-forcing trajectory regression limits trajectory diversity.
- Slightly slower inference (7.8 FPS vs 9.0 FPS for SparseDrive).
- Turning-nuScenes is a small curated validation subset (17 scenes, 680 samples) and does not appear to be released.
Experimental setup
Datasets: nuScenes Bench2Drive
Benchmarks: nuScenes Turning-nuScenes Bench2Drive
Metrics: L2 error collision rate TPC mAP NDS AMOTA driving score success rate comfort
Baselines: UniAD VAD SparseDrive VADv2 VAD-mmt
Builds on: SparseDrive VAD UniAD VADv2
Closest papers by meaning
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.88)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.87)
- DualAD: Disentangling the Dynamic and Static World for End-to-End Driving (CVPR 2024 · similarity 0.86)
- $AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.86)
- AlignDrive: Aligned Lateral-Longitudinal Planning for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.86)
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · 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.