SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving
SparseDriveV2 leverages dense static vocabularies with trajectory decomposition, achieving 92.0 PDMS, challenging the necessity of dynamic proposals.
Key Findings
Methodology
This work systematically investigates the impact of trajectory vocabulary density using Hydra-MDP as a baseline. It reveals performance improvements with increased anchor density, without saturation under computational limits. SparseDriveV2 introduces a factorized trajectory representation—geometric paths and velocity profiles—and employs a hierarchical scoring strategy: coarse scoring over components followed by fine-grained evaluation of composed trajectories. This approach enables a 32× increase in vocabulary density while maintaining computational efficiency, validated on NAVSIM and Bench2Drive benchmarks, surpassing state-of-the-art methods.
Key Results
- On NAVSIM, SparseDriveV2 achieves 92.0 PDMS and 90.1 EPDMS, outperforming existing static vocabulary methods; on Bench2Drive, it scores 89.15 Driving Score and 70.00 success rate, demonstrating the effectiveness of ultra-dense vocabularies combined with hierarchical scoring.
- Incremental vocabulary expansion shows continuous performance gains, indicating that static vocabularies, when sufficiently dense, can match dynamic proposal methods.
- The factorized structure reduces model complexity and enhances scalability, with experiments confirming efficiency and robustness across diverse scenarios.
Significance
This research challenges the prevailing view that dynamic trajectory generation is essential for high-performance end-to-end autonomous driving. By demonstrating that a densely populated static vocabulary, combined with an efficient multi-stage scoring pipeline, can achieve or surpass state-of-the-art results, it opens new avenues for scalable, computationally efficient planning systems. This approach simplifies model architecture, reduces reliance on complex generative modules, and provides a practical pathway for deploying autonomous vehicles in real-world environments with limited computational resources.
Technical Contribution
The core innovation lies in the trajectory factorization into geometric paths and velocity profiles, enabling combinatorial coverage of the action space. The hierarchical scoring mechanism—initial coarse filtering followed by detailed evaluation—significantly reduces computational load while maintaining accuracy. The large-scale data-driven clustering of paths and velocities constructs a super-dense yet compact vocabulary, enabling the model to explore a richer set of candidate trajectories without exponential growth in complexity. This framework enhances scalability and paves the way for future large-scale, high-density planning systems.
Novelty
This is the first comprehensive study to demonstrate that static trajectory vocabularies, when densely populated through factorization and hierarchical scoring, can rival or outperform dynamic proposal methods. The key novelty is the combination of trajectory decomposition with multi-stage scoring, which effectively expands the action space coverage without increasing model complexity prohibitively. This challenges the assumption that dynamic generation is indispensable, offering a new paradigm for scalable end-to-end planning.
Limitations
- While the dense static vocabulary improves coverage, it may still struggle in highly dynamic or unpredictable environments where real-time adaptation is critical. The approach relies on precomputed clusters, which may not generalize well to unseen scenarios.
- Despite efficiency gains, the hierarchical scoring process still incurs computational costs, especially in extremely dense vocabularies, limiting real-time deployment in resource-constrained systems.
- The current validation is primarily in simulation; real-world factors such as sensor noise, environmental variability, and unexpected obstacles require further investigation to ensure robustness.
Future Work
Future research will focus on adaptive vocabulary updating, integrating online learning to refine trajectory sets based on real-time data. Combining multi-modal sensor inputs, such as LiDAR and semantic maps, could further improve coverage and accuracy. Additionally, extending this framework to multi-agent scenarios and multi-task learning could broaden its applicability, pushing autonomous driving toward more scalable and resilient solutions.
AI Executive Summary
Autonomous driving systems aim to navigate complex traffic environments efficiently and safely. Traditional approaches often rely on either static trajectory vocabularies—discrete sets of pre-planned paths—or dynamic proposal generation, which involves complex generative models. Static vocabularies are limited by coarse discretization, while dynamic methods, such as diffusion models or regression-based proposals, are computationally intensive and complex to deploy at scale.
This paper introduces SparseDriveV2, a novel framework that challenges the necessity of dynamic proposals by leveraging a dense static vocabulary constructed through trajectory decomposition. By splitting trajectories into geometric paths and velocity profiles, the authors enable combinatorial coverage of the action space, significantly increasing the number of candidate trajectories—up to 32 times more than prior static methods. A hierarchical scoring pipeline first performs coarse evaluations over paths and velocities independently, then fine-grained scoring on a small subset of combined trajectories, ensuring computational efficiency.
Experimental results on NAVSIM and Bench2Drive benchmarks demonstrate that SparseDriveV2 achieves state-of-the-art performance, with a PDMS of 92.0 and a Driving Score of 89.15, outperforming existing static and dynamic methods. These findings suggest that with sufficient density and effective scoring, static vocabularies can match or surpass the performance of more complex dynamic proposal systems. This work opens new avenues for scalable, efficient, and high-performance autonomous planning, reducing model complexity while maintaining robustness.
Despite these advances, challenges remain in adapting to highly unpredictable real-world environments and further optimizing computational costs for real-time deployment. Future directions include online vocabulary adaptation, multi-modal sensor fusion, and broader multi-agent applications, aiming to bring autonomous driving closer to practical, widespread use.
Deep Dive
Plain Language Accessible to non-experts
想象你在准备一份超级丰富的菜单，菜单上有成千上万的菜肴选择。传统的方法就像提前准备好一份有限的菜单，只包含几百个菜肴，虽然简单，但在面对复杂的宴会时，菜单可能不够丰富，不能满足所有客人的需求。而这篇论文提出的方法，就像是把菜肴拆分成不同的部分，比如主料和调料，然后用不同的组合方式，创造出无数新菜肴。这样一来，即使菜单变得超级丰富，也不用增加太多的准备工作。系统会先快速筛选出一些看起来不错的菜，然后再详细品尝，确保每一道菜都符合口味。最终，这样的菜单既丰富又高效，能满足各种复杂的宴会需求。这就像是用聪明的拆分和筛选，让菜单变得既丰富又实用，帮助厨师做出最棒的菜肴。
ELI14 Explained like you're 14
想象你在玩一个超级复杂的拼图游戏，拼图块有成千上万种不同的形状和颜色。以前的方法就像是提前准备好一份有限的拼图，虽然简单，但拼不出很多复杂的图案。这篇论文的方法就像是把拼图拆成两个部分：一部分是拼图的基本形状，另一部分是颜色和细节。然后用一种聪明的办法，把这些部分组合起来，拼出很多不同的图案。游戏开始时，系统会先快速挑出一些看起来不错的拼图组合，然后再仔细检查，确保拼得漂亮。这样一来，即使拼图很多，也不用花太多时间，就能拼出漂亮的图案。这就像是用聪明的拆分和筛选，让拼图变得既丰富又容易拼，帮你在游戏中取得好成绩。
Abstract
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.