[CVPR2025] Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving

Abstract

End-to-end autonomous driving frameworks facilitate seamless integration of perception and planning but often rely on one-shot trajectory prediction, lacking temporal consistency and long-horizon awareness. This limitation can lead to unstable control, undesirable shifts, and vulnerability to occlusions in single-frame perception. In this work, we propose the Momentum-Aware Driving (MomAD) framework to address these issues by introducing trajectory momentum and perception momentum to stabilize and refine trajectory prediction. MomAD consists of two key components: (1) Topological Trajectory Matching (TTM), which uses Hausdorff Distance to align predictions with prior paths and ensure temporal coherence, and (2) Momentum Planning Interactor (MPI), which cross-attends the planning query with historical spatial-temporal context. Additionally, an encoder-decoder module introduces feature perturbations to increase robustness against perception noise. To quantify planning stability, we propose the Trajectory Prediction Consistency (TPC) metric, showing that MomAD achieves long-term consistency (>3s) on the nuScenes dataset. We further curate the challenging Turning-nuScenes validation set, focused on turning scenarios, where MomAD surpasses state-of-the-art methods, highlighting its enhanced stability and responsiveness in dynamic driving conditions.

🔥 Contributions:

• Momentum Planning Concept. We propose the concept of momentum planning in multi-modal trajectory planning, drawing an analogy to human driving behavior. We provide theoretical evidence to demonstrate the effectiveness of our momentum planning in addressing temporal consistency in end-to-end autonomous driving

• MomAD Framework. We propose MomAD, an end-to-end autonomous driving framework that employs momentum planning. It optimizes current trajectory planning by integrating historical planning guidance, significantly improving trajectory consistency and stability in autonomous driving.

• Turning NuScenes Validation Dataset. We create the Turning-nuScenes val dataset, derived from the nuScenes full validation dataset. This new dataset focuses on turning scenarios, providing a specialized benchmark for evaluating the performance of autonomous driving systems in complex driving situations.

• Trajectory Prediction Consistency (TPC) Metric. We introduce the TPC metric to quantitatively assess the consistency of trajectory predictions in existing end-to-end autonomous driving methods, addressing a critical gap in the evaluation of trajectory planning.

• Performance Evaluation. Through extensive experiments on the nuScenes dataset, we demonstrate that MomAD significantly outperforms SOTA methods in terms of trajectory consistency and stability, highlighting its effectiveness in tackling challenges within autonomous driving planning. We evaluated the results of long trajectory predictions, specifically at 4, 5, and 6 seconds, which are critical for ensuring the stability of autonomous driving systems.

Method

The overall architecture of MomAD. MomAD, as a multi-model trajectory end-to-end autonomous driving method, first encodes multi-view images into feature maps, then learns a sparse scene representation through sparse perception, and finally performs a momentum-guided motion planner to accomplish the planning task. The momentum planning module integrates historical planning to inform current planning, effectively addressing the issue of maximum score deviation in multi-modal trajectories.

Results in paper

Open-loop mertics

• Planning results on nuScenes.

Method	Backbone	L2 (m) 1s	L2 (m) 2s	L2 (m) 3s	L2 (m) Avg	Col. (%) 1s	Col. (%) 2s	Col. (%) 3s	Col. (%) Avg	TPC (m) 1s	TPC (m) 2s	TPC (m) 3s	TPC (m) Avg	FPS
UniAD	ResNet101	0.45	0.70	1.04	0.73	0.62	0.58	0.63	0.61	0.41	0.68	0.97	0.68	1.8 (A100)
SparseDrive	ResNet50	0.29	0.58	0.96	0.61	0.01	0.05	0.18	0.08	0.30	0.57	0.85	0.57	9.0 (4090)
MomAD (Ours)	ResNet50	0.31	0.57	0.91	0.60	0.01	0.05	0.22	0.09	0.30	0.53	0.78	0.54	7.8 (4090)

• Planning results for long trajectory prediction on nuScenes. We train 10 epochs on 6s trajectories and test on 6s trajectories.

Method	L2 (m) 4s	L2 (m) 5s	L2 (m) 6s	Col. (%) 4s	Col. (%) 5s	Col. (%) 6s	TPC (m) 4s	TPC (m) 5s	TPC (m) 6s
SparseDrive	1.75	2.32	2.95	0.87	1.54	2.33	1.33	1.66	1.99
MomAD (Ours)	1.67	1.98	2.45	0.83	1.43	2.13	1.19	1.45	1.61

• Planning results on the Turning-nuScenes validation dataset Turning-nuScenes . We train 10 epochs on 6s trajectories and test on 6s trajectories.

Method	L2 (m) 1s	L2 (m) 2s	L2 (m) 3s	Col. (%) 1s	Col. (%) 2s	Col. (%) 3s	TPC (m) 1s	TPC (m) 2s	TPC (m) 3s
SparseDrive	0.35	0.77	1.46	0.86	0.04	0.17	0.98	0.40	0.34
MomAD (Ours)	0.33	0.70	1.24	0.76	0.03	0.13	0.79	0.32	0.32

Close-loop mertics

• Open-loop and Closed-loop Results of E2E-AD Methods in Bench2Drive (V0.0.3)} under base training set. `mmt' denotes the extension of VAD on Multi-modal Trajectory. * denotes our re-implementation. The metircs momad used follows Bench2Drive

Method	Open-loop Metric	Closed-loop Metric
	Avg. L2 ↓	DS ↑	SR(%) ↑	Effi ↑	Comf ↑
VAD	0.91	42.35	15.00	157.94	46.01
VAD mmt*	0.89	42.87	15.91	158.12	47.22
Our MomAD (Euclidean)	0.84	46.12	17.45	173.35	50.98
Our MomAD	0.85	45.35	17.44	162.09	49.34
SparcDrive*	0.87	44.54	16.71	170.21	48.63
Our MomAD (Euclidean)	0.84	46.12	17.45	173.35	50.98
Our MomAD	0.82	47.91	18.11	174.91	51.20

Close_loop Vis

Robustness evaluation

• Robustness analysis on nuScenes-C

Setting	Method	Detection		Tracking	Mapping	Motion	Planning
		mAP ↑	NDS ↑	AMOTA ↑	mAP ↓	mADE ↓	L2 ↓	Col. ↓	TPC ↓
Clean	SparseDrive	0.418	0.525	0.386	55.1	0.62	0.61	0.08	0.57
Clean	Our MomAD	0.423	0.531	0.391	55.9	0.61	0.60	0.09	0.54
Snow	SparseDrive	0.091	0.111	0.102	16.0	0.98	0.88	0.32	0.82
Snow	Our MomAD	0.154	0.173	0.166	20.9	0.76	0.73	0.16	0.68
Fog	SparseDrive	0.141	0.159	0.154	18.8	0.91	0.86	0.41	0.80
Fog	Our MomAD	0.197	0.197	0.206	24.9	0.73	0.71	0.18	0.67
Rain	SparseDrive	0.128	0.140	0.193	19.4	0.97	0.93	0.46	0.92
Rain	Our MomAD	0.207	0.213	0.266	25.2	0.76	0.71	0.21	0.71

Trajectory Prediction Consistency (TPC) metric

To evaluate the planning stability of MomAD, we propose a new Trajectory Prediction Consistency (TPC) metric to measure consistency between predicted and historical trajectories.

How to generate a 6s nuScenes trajectory dataset？

python tools/data_converter/nuscenes_converter_6s.py nuscenes \
    --root-path ./data/nuscenes \
    --canbus ./data/nuscenes \
    --out-dir ./data/infos/ \
    --extra-tag nuscenes \
    --version v1.0

Quick Start

Quick Start for Open_loop

Quick start for Close_loop

Acknowledgement

• SparseDrive
• UniAD
• VAD
• mmdet3d