SparseDriveV2: Scalable Score-Based Planning

• The paper introduces a factorized trajectory vocabulary and a hierarchical scoring pipeline that leverages separate path and velocity representations to achieve state-of-the-art planning performance.
• It scales the static vocabulary by decomposing trajectories into geometric paths and velocity profiles, enabling a 32× denser candidate set while mitigating memory and computation issues.
• Empirical evaluations on benchmarks like NAVSIM and Bench2Drive demonstrate significant improvements in PDMS, EPDMS, and Driving Score, validating the method's efficiency with a lightweight ResNet-34 backbone.

Searching arXiv for the specified paper and closely related planning methods to ground the article. {"query":"SparseDriveV2 Scoring is All You Need for End-to-End Autonomous Driving arXiv (Sun et al., 31 Mar 2026)", "max_results": 5} {"query":"Hydra-MDP end-to-end autonomous driving arXiv", "max_results": 10} SparseDriveV2 is an end-to-end autonomous driving planner organized around scoring-based multimodal planning rather than direct single-trajectory regression. The method enumerates many plausible candidate trajectories, assigns each a score, and selects the highest-scoring trajectory for control, treating multimodality as a first-class property of driving scenes such as intersections, merges, lane changes, and interactions around stopped vehicles. The central question posed by the work is whether strong end-to-end planning performance truly requires dynamic proposal generation, or whether a sufficiently dense static vocabulary can already cover the action space effectively if it is represented and scored in a scalable manner. SparseDriveV2 answers this question with a factorized trajectory vocabulary and a hierarchical scoring pipeline, and reports 92.0 PDMS and 90.1 EPDMS on NAVSIM, together with 89.15 Driving Score and 70.00 Success Rate on Bench2Drive using a lightweight ResNet-34 backbone (Sun et al., 31 Mar 2026).

1. Problem formulation and research context

End-to-end multi-modal planning commonly models uncertainty by scoring candidate trajectories and selecting the optimal one. Within this family, prior approaches are described as belonging to two broad groups. Static-vocabulary methods precompute a finite set of anchors, often by clustering trajectories from data; the paper lists VADv2, Hydra-MDP, Hydra-MDP++, and DriveSuprim as representative examples. These methods are simple and efficient at inference, but the vocabulary is typically only a few thousand trajectories, so the action space is discretized too coarsely. Dynamic-proposal methods, including regression-based, diffusion-based, or flow-based planners, can generate finer-grained candidates conditioned on the scene and often outperform static methods on benchmarks, but they introduce extra model components, iterative sampling or denoising, or more complex training (Sun et al., 31 Mar 2026).

The conceptual intervention of SparseDriveV2 is to revisit the assumption that dynamic generation is fundamentally necessary. The work argues that the apparent advantage of dynamic methods may stem less from generation itself than from better action-space coverage. This suggests that the central systems question is not simply whether trajectories are generated online, but whether the planner can maintain dense, scene-relevant coverage while preserving tractable scoring and inference.

SparseDriveV2 therefore remains within a purely scoring-based paradigm. Its two coupled design choices are a factorized vocabulary representation and a scalable scoring strategy. Together, these are intended to recover the coverage advantages usually associated with dynamic proposal generation without resorting to diffusion or iterative proposal synthesis (Sun et al., 31 Mar 2026).

2. Scaling evidence from Hydra-MDP

The paper’s first empirical contribution is a scaling study of Hydra-MDP, presented as a representative static-vocabulary scorer. The study increases the number of trajectory anchors and measures both performance and memory consumption. The reported trend is monotonic: as anchor count rises from 1024 to 16384, EPDMS on NAVSIM v2 increases from 85.02 to 87.35, with consistent gains at every scale and no saturation before computational constraints become prohibitive (Sun et al., 31 Mar 2026).

Anchor count	EPDMS on NAVSIM v2	Memory
1024	85.02	9.5 GB
16384	87.35	38.9 GB
32768	out of memory	out of memory

The significance of this scaling result is methodological. It supports the thesis that static vocabularies are not inherently weak; rather, naïve scaling of a monolithic static vocabulary is computationally expensive. A plausible implication is that prior performance gaps between static and dynamic planners may reflect representational and systems bottlenecks more than an intrinsic deficiency of scoring-based planning itself.

This scaling study directly motivates SparseDriveV2. If denser anchor sets continue to improve performance without saturation, then the critical problem becomes how to enlarge effective coverage without incurring the quadratic or memory-heavy costs associated with scoring a very large flat trajectory bank.

3. Factorized trajectory vocabulary

SparseDriveV2 addresses vocabulary scaling by factorizing a trajectory into two components: a geometric path and a velocity profile. A trajectory is defined as

$\tau = \{(x_t, y_t)\}_{t=1}^{T},$

with fixed temporal sampling. The factorization is written as

$p = \{(x_i, y_i)\}_{i=1}^{S}$

for the path, sampled at fixed spatial interval $\Delta s$, and

$v = \{v_t\}_{t=1}^{T}$

for the velocity profile, sampled at fixed temporal interval $\Delta t$ (Sun et al., 31 Mar 2026).

Extraction from a trajectory is expressed as

$(p, v) = \mathcal{D}(\tau),$

where the path is obtained by resampling the spatial shape of the trajectory along traveled distance, and the velocity is computed by

$$v_t = \frac{\| (x_t, y_t) - (x_{t-1}, y_{t-1}) \|}{\Delta t}.$$

Reconstruction is performed by composition,

$\tau = \mathcal{C}(p, v),$

with cumulative traveled distance

$s_t = \sum_{k=1}^{t} v_k \Delta t,$

and the position at each time recovered by interpolating along the path at distance $s_t$ (Sun et al., 31 Mar 2026).

The scalability gain comes from clustering paths and velocities separately:

$p = \{(x_i, y_i)\}_{i=1}^{S}$0

and constructing the full trajectory vocabulary by Cartesian composition:

$p = \{(x_i, y_i)\}_{i=1}^{S}$1

so that

$p = \{(x_i, y_i)\}_{i=1}^{S}$2

In the main NAVSIM configuration, the model uses 1024 paths and 256 velocities, yielding 262,144 trajectories, which the paper describes as 32× denser than the common 8192-anchor static vocabularies used by prior methods (Sun et al., 31 Mar 2026).

The underlying structural claim is that many useful trajectories share geometry but differ in speed, or share a speed profile but differ in geometry. Factorization exploits this regularity rather than learning each full trajectory independently. This suggests a more general principle for end-to-end planning: decompositions that preserve semantically meaningful independence structure can substantially expand action-space coverage without proportionally increasing the size of the learned primitive set.

4. Hierarchical scoring architecture

A dense combinatorial vocabulary is only useful if it can be scored efficiently. SparseDriveV2 therefore uses a two-stage scoring pipeline consisting of coarse factorized scoring followed by fine-grained scoring over composed trajectories (Sun et al., 31 Mar 2026).

In the coarse stage, given scene features

$p = \{(x_i, y_i)\}_{i=1}^{S}$3

the model encodes each path anchor and velocity anchor into embeddings:

$p = \{(x_i, y_i)\}_{i=1}^{S}$4

Each path embedding is then contextually refined by path-scene interaction,

$p = \{(x_i, y_i)\}_{i=1}^{S}$5

and scored as

$p = \{(x_i, y_i)\}_{i=1}^{S}$6

Velocity profiles are processed analogously:

$p = \{(x_i, y_i)\}_{i=1}^{S}$7

For paths, the interaction may be implemented with multi-head cross-attention or deformable aggregation along the path geometry. Functionally, this stage acts as a fast filter that removes implausible components early, such as high-speed velocities in congestion or turning paths on a straight road (Sun et al., 31 Mar 2026).

After factorized scoring, the model retains only the top-$p = \{(x_i, y_i)\}_{i=1}^{S}$8 paths and top-$p = \{(x_i, y_i)\}_{i=1}^{S}$9 velocities, forming a reduced candidate set

$\Delta s$0

whose size is $\Delta s$1, far smaller than the full Cartesian vocabulary.

The fine stage reintroduces joint reasoning at the composed-trajectory level. For each candidate $\Delta s$2, the model fuses path and velocity representations as

$\Delta s$3

then applies trajectory-scene re-conditioning,

$\Delta s$4

followed by final scoring,

$\Delta s$5

This second-stage re-conditioning is necessary because path and velocity are not truly independent. A sharp curve at high speed may be unsafe even when the path and velocity are individually plausible. The architecture therefore performs cheap factorized reasoning broadly and expensive joint spatiotemporal reasoning narrowly, which is the paper’s explicit solution to the scalability problem (Sun et al., 31 Mar 2026).

5. Training objective and inference

SparseDriveV2 is trained as a scoring model using soft targets derived from distances to expert behavior. For paths, the masked squared distance to the ground-truth geometric path $\Delta s$6 is

$\Delta s$7

and path scoring is supervised with

$\Delta s$8

For velocity profiles,

$\Delta s$9

with loss

$v = \{v_t\}_{t=1}^{T}$0

For composed trajectories,

$v = \{v_t\}_{t=1}^{T}$1

and

$v = \{v_t\}_{t=1}^{T}$2

The model also uses metric supervision distilled from a rule-based teacher, with binary cross-entropy over sub-scores such as safety, progress, comfort, and rule compliance:

$v = \{v_t\}_{t=1}^{T}$3

The total objective is

$v = \{v_t\}_{t=1}^{T}$4

At inference time, the planner aggregates predicted sub-scores using the benchmark’s evaluation weighting and returns the trajectory with the highest final score (Sun et al., 31 Mar 2026).

This training design is notable because it combines imitation-oriented proximity supervision with metric-oriented teacher supervision. A plausible implication is that the model is not merely ranking trajectories by geometric similarity to expert data, but also learning benchmark-aligned notions of safety, comfort, progress, and rule compliance as explicit scoring dimensions.

6. Experimental protocol, results, and interpretation

The evaluation spans NAVSIM and Bench2Drive. NAVSIM v1 is described as an open-loop benchmark with metrics NC, DAC, TTC, Comfort, and EP, aggregated into PDMS. NAVSIM v2 extends the metric set to NC, DAC, DDC, TL, EP, TTC, LK, HC, EC, aggregated into EPDMS. Bench2Drive is a closed-loop CARLA benchmark with Driving Score, Success Rate, Driving Efficiency, and Comfort, together with multi-ability scores (Sun et al., 31 Mar 2026).

Implementation on NAVSIM uses a ResNet-34 image backbone and selected front-camera images at $v = \{v_t\}_{t=1}^{T}$5 resolution. The core vocabulary consists of 1024 path anchors and 256 velocity anchors. For efficient inference, the system first retains top 128 paths and 64 velocities, then refines to 20 paths and 20 velocities, yielding 400 final trajectory hypotheses for fine-grained scoring. On NAVSIM v2, the second-stage velocity count is reduced to 10 for faster metric computation. Training is performed on 8 L20 GPUs for 10 epochs (Sun et al., 31 Mar 2026).

Benchmark	Metric	Result
NAVSIM v1	PDMS	92.0
NAVSIM v2	EPDMS	90.1
Bench2Drive	Driving Score	89.15
Bench2Drive	Success Rate	70.00
Bench2Drive	Multi-ability score	67.67

The paper states that 92.0 PDMS on NAVSIM v1 is the best reported result in its comparison table, outperforming all listed scoring-based and dynamic-generation baselines. On NAVSIM v2, SparseDriveV2 reaches 90.1 EPDMS under the corrected official evaluation, surpassing DiffusionDriveV2 by 2.6 points and outperforming strong scoring-based baselines with larger backbones. On Bench2Drive, it reports 89.15 Driving Score and 70.00% Success Rate, with a 67.67 multi-ability score, again presented as a new high among the compared methods (Sun et al., 31 Mar 2026).

The ablation studies are aligned with the paper’s two main claims. First, increasing $v = \{v_t\}_{t=1}^{T}$6 and $v = \{v_t\}_{t=1}^{T}$7 improves EPDMS monotonically; for example, moving from 512×128 to 1024×256 improves EPDMS from 88.7 to 90.1. Second, deformable aggregation performs better than plain cross-attention in the path-scene interaction stage. Third, adding trajectory re-conditioning consistently improves results for both interaction types, indicating that independent path and velocity scoring is insufficient without a final joint compatibility model (Sun et al., 31 Mar 2026).

The broader interpretation offered by the work is that scoring-based planning is not fundamentally limited by static vocabularies. If the vocabulary is represented factorwise and scored hierarchically, a purely scoring-based planner can match or exceed the performance of more complex dynamic-generation approaches. The stated limitation is that the model can still fail when navigation information is insufficient, leading to incorrect high-level route decisions in some scenarios. More generally, the framework remains dependent on the quality of the factorized vocabulary and on the teacher or metric supervision used for scoring. Even with those caveats, the reported evidence supports a clear thesis: in end-to-end autonomous driving, better coverage plus better scoring may close much of the gap to dynamic proposal generation while preserving a simpler and more interpretable planning pipeline (Sun et al., 31 Mar 2026).