Source-linked AI summary
SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving
Wenchao Sun, Xuewu Lin, Keyu Chen, Zixiang Pei, Xiang Li, Yining Shi, Sifa Zheng
TL;DR
High-performance end-to-end planning may not require dynamic trajectory generation if static vocabularies are sufficiently dense and efficiently scored. SparseDriveV2 factorizes trajectories into paths and velocity profiles, combines coarse factorized scoring with fine-grained scoring, and reports state-of-the-art results across NAVSIM and Bench2Drive. Its formulation also recognizes that path–velocity independence does not always hold and introduces trajectory re-conditioning for joint reasoning.
Problem
It remains unclear whether dynamic trajectory generation is necessary for high-performance planning or whether sufficiently dense static vocabularies can achieve comparable performance.
Method
SparseDriveV2 factorizes trajectories into geometric paths and velocity profiles, then uses coarse factorized scoring followed by fine-grained scoring on composed trajectories.
Results
SparseDriveV2 achieves state-of-the-art performance among scoring-based and dynamic-generation methods on NAVSIM and Bench2Drive.
Takeaways & Limitations
Sufficiently dense vocabularies and efficient scoring mechanisms can make scoring-based planning a simple yet powerful paradigm for end-to-end autonomous driving.
Takeaways & Limitations
Path–velocity independence does not always hold in real-world driving, so trajectory re-conditioning is introduced to capture spatiotemporal dependencies.
Abstract
from arXiv · showhide
End-to-end multi-modal planning has been widely adopted to model the uncertainty of driving behavior, typically by scoring candidate trajectories and selecting the optimal one. Existing approaches generally fall into two categories: scoring a large static trajectory vocabulary, or scoring a small set of dynamically generated proposals. While static vocabularies often suffer from coarse discretization of the action space, dynamic proposals provide finer-grained precision and have shown stronger empirical performance on existing benchmarks. However, it remains unclear whether dynamic generation is fundamentally necessary, or whether static vocabularies can already achieve comparable performance when they are sufficiently dense to cover the action space. In this work, we start with a systematic scaling study of Hydra-MDP, a representative scoring-based method, revealing that performance consistently improves as trajectory anchors become denser, without exhibiting saturation before computational constraints are reached. Motivated by this observation, we propose SparseDriveV2 to push the performance boundary of scoring-based planning through two complementary innovations: (1) a scalable vocabulary representation with a factorized structure that decomposes trajectories into geometric paths and velocity profiles, enabling combinatorial coverage of the action space, and (2) a scalable scoring strategy with coarse factorized scoring over paths and velocity profiles followed by fine-grained scoring on a small set of composed trajectories. By combining these two techniques, SparseDriveV2 achieves 92.0 PDMS and 90.1 EPDMS on NAVSIM, with 89.15 Driving Score and 70.00 Success Rate on Bench2Drive with a lightweight ResNet-34 as backbone. Code and model are released at https://github.com/swc-17/SparseDriveV2.
1 Introduction
The paper asks whether dense static vocabularies can match dynamic trajectory generation and addresses this through scaling and factorized scoring. SparseDriveV2 combines compact combinatorial coverage with efficient refinement, achieving state-of-the-art benchmark performance.
- Motivation: Dynamic proposals offer finer-grained coverage than static vocabularies, but their necessity for high-performance planning remains unresolved.Static methods are constrained by coarse action-space discretization, while dynamic generation has shown stronger benchmark performance.
- Scaling study: Hydra-MDP performance consistently improves as trajectory anchors become denser, without saturation before computational constraints are reached.This suggests static-vocabulary limitations stem from insufficient action-space coverage rather than an intrinsic restriction.
- Method: SparseDriveV2 factorizes trajectories into geometric paths and velocity profiles, enabling combinatorial coverage with a compact vocabulary.The representation addresses the challenge of representing extremely dense candidate sets under memory and computation limits.
- Method: Its scoring strategy first prunes candidates through coarse factorized scoring, then applies fine-grained scoring to a small set of composed trajectories.Precise spatiotemporal reasoning is restricted to high-quality candidates rather than the full vocabulary.
- Results: 32× denser vocabulary enables SparseDriveV2 to achieve state-of-the-art performance on NAVSIM and Bench2Drive.The vocabulary scales from 8192 prior anchors to 1024×256 composed candidates while retaining efficient scoring.
2 Related Works
Prior work uses either predefined trajectory vocabularies or dynamically generated proposals, each addressing multimodal planning with different coverage and computational trade-offs. SparseDriveV2 instead remains purely scoring-based while using factorized trajectory construction to expand coverage.
- Scoring-based methods: Scoring-based methods select an optimal trajectory from a predefined candidate set, but vocabulary size makes finer discretization computationally expensive.Examples include VADv2, Hydra-MDP, and DriveSuprim.
- Dynamic generation: Dynamic generation conditions trajectory proposals on scene information to provide more flexible and expressive candidates.Regression- and diffusion-based approaches support finer-grained adaptation and broader action-space coverage.
- Hybrid methods: GTRS combines diffusion-based dynamic proposals with static candidates to address complementary weaknesses in coverage and generalization.Its motivation reflects static vocabularies’ coarse discretization and small dynamic sets’ limited coverage of unseen trajectories.
- SparseDriveV2: SparseDriveV2 follows a purely scoring-based paradigm without dynamic trajectory generation.Its factorized vocabulary decomposes trajectories into spatial and temporal components for compact combinatorial coverage.
3 Methods
SparseDriveV2 factorizes trajectories into geometric paths and velocity profiles, composes them combinatorially, and scores the resulting candidates efficiently through coarse component-level filtering and trajectory-level scoring.
- Problem Formulation: Scoring-based planning evaluates candidate trajectories and selects the one with the highest score for vehicle control.The scorer assigns each candidate a score based on scenario-relevant criteria.
- Scalable Vocabulary Representation: A trajectory is factorized into a geometric path describing spatial shape and a velocity profile describing temporal speed.The path contains no temporal information, while the velocity profile specifies progression speed independently of spatial geometry.
- Scalable Vocabulary Representation: The factorized representation expands trajectory coverage while preserving a compact vocabulary structure.The resulting trajectories are reconstructed by composing a path with a velocity profile.
- Scalable Vocabulary Construction: Path and velocity vocabularies are composed pairwise to form a trajectory vocabulary whose size scales as |T| = Np × Nv.This combinatorial construction provides a super-dense trajectory set from compact component vocabularies.
- Scalable Scoring: Coarse path and velocity scores select top-K components before their combinations undergo trajectory-level scoring.Independent component scoring filters implausible combinations, while joint trajectory features support final scoring.
4 Experiments
Experiments evaluate SparseDriveV2 on NAVSIM and Bench2Drive, including its dense factorized vocabulary, progressive filtering, and benchmark performance. The method uses 1024 path anchors and 256 velocity anchors to form 262,144 trajectory anchors while scoring a smaller filtered set.
- Datasets and metrics: NAVSIM v1 aggregates no collision, drivable area compliance, ego progress, time-to-collision, and comfort into PDMS.NAVSIM v2 adds driving direction compliance, traffic light compliance, lane keeping, and extended comfort for EPDMS.
- Datasets and metrics: Bench2Drive evaluates closed-loop planning across 220 test routes and 44 interactive scenarios using Driving Score, Success Rate, Efficiency, and Comfortness.
- Implementation details: 262,144 trajectory anchors result from composing 1024 path anchors with 256 velocity anchors.This factorized vocabulary is 32× denser than the commonly used 8192-anchor vocabulary in prior scoring-based methods.
- Implementation details: 400 trajectory hypotheses remain for fine-grained scoring after progressive filtering retains 128 paths and 64 velocities, then refines them to 20 paths and 20 velocities.Metric supervision is applied only to these 400 filtered trajectories.
- Results: 92.0 PDMS is achieved by SparseDriveV2 on NAVSIM v1 with a 21.8M-parameter ResNet-34 backbone.The method also achieves 90.1 EPDMS on NAVSIM v2 and surpasses DiffusionDriveV2 by 2.6 EPDMS under the corrected metric.
- Ablations: Increasing vocabulary size consistently improves planning performance in the NAVSIM v2 ablation.The ablation evaluates scaling path and velocity anchors alongside the scalable scoring design.
5 Conclusion
SparseDriveV2 introduces scalable vocabulary and scoring mechanisms for super-dense candidate trajectories. Experiments report state-of-the-art performance across scoring-based and dynamic-generation methods, supporting dense scoring-based planning as a powerful paradigm.
- Contributions: SparseDriveV2 introduces a scalable trajectory vocabulary representation and scalable scoring strategy for super-dense candidate trajectories.
- Ablations: Ablation studies evaluate factorized vocabulary scaling and scalable scoring through path–scene interaction and trajectory re-conditioning.The scoring comparison includes multi-head cross-attention and deformable aggregation.
- Conclusion: SparseDriveV2 achieves state-of-the-art performance among scoring-based and dynamic-generation methods.The conclusion attributes this result to sufficiently dense vocabularies combined with efficient scoring mechanisms.
A Implementation Details for Bench2Drive
Bench2Drive uses a factorized vocabulary, multi-camera ResNet-50 scene encoding, and two-stage training with imitation learning. Selected trajectories are decomposed into paths and velocity profiles for vehicle control.
- Vocabulary: The Bench2Drive vocabulary contains 1024 path anchors and 256 velocity anchors sampled over 15 m and 3 s horizons.
- Scene encoding: A ResNet-50 processes six camera views, while object detection, online mapping, and motion prediction provide auxiliary tasks.Path and velocity embeddings interact with agent and map queries.
- Training: Training first optimizes perception tasks for 100 epochs, then jointly trains planning and perception for 10 additional epochs.Bench2Drive uses pure imitation learning without metric-based supervision.
- Vehicle control: After scoring, the selected trajectory is decomposed into a geometric path and velocity profile for control.The path supplies lateral-control preview geometry, while the first velocity supplies the target speed for throttle and braking.
- Vehicle control: The lateral-control preview distance is d = 0.5 × ego_speed + 2.5.
B Qualitative Results
Qualitative results compare SparseDriveV2 with expert and baseline trajectories, highlighting smoother turning, higher traffic efficiency, and improved intent alignment. Failure cases show incorrect navigation decisions that may reflect insufficient navigation information.
- Visualization: NAVSIM visualizations distinguish the human-expert ground truth, GTRS-Dense baseline, and SparseDriveV2 trajectories by color.The expert, baseline, and proposed trajectories are shown in green, red, and blue, respectively.
- Trajectory quality: SparseDriveV2 produces smoother trajectories than the baseline in sharp-turning scenarios.
- Traffic efficiency: SparseDriveV2 achieves higher traffic efficiency while the baseline remains stationary.
- Intent alignment: SparseDriveV2 aligns more closely with the expert trajectory in high-level intent through geometric path modeling.
- Failure cases: Failure cases include incorrect navigation decisions, possibly due to insufficient navigation information.