Necessity of Dynamic Trajectory Generation vs. Sufficiently Dense Static Vocabularies

Determine whether dynamic trajectory generation methods for end-to-end autonomous driving planning are fundamentally necessary for high-performance planning, or whether static trajectory vocabularies—when scaled to be sufficiently dense to cover the action space and paired with effective scoring—can achieve comparable performance.

Background

End-to-end autonomous driving planners often adopt multi-modal planning that scores candidate trajectories, with two main paradigms: scoring a static trajectory vocabulary or generating dynamic proposals. Dynamic generation methods (e.g., regression- or diffusion-based) have shown strong empirical performance but introduce additional complexity.

The paper investigates whether dynamic proposal generation is inherently required, or if a sufficiently dense static trajectory vocabulary, combined with an efficient scoring mechanism, can already achieve comparable outcomes. This uncertainty motivates their scaling study and the design of a factorized trajectory vocabulary and scalable scoring in SparseDriveV2.

References

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.

— SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving  (2603.29163 - Sun et al., 31 Mar 2026) in Abstract

The open question is how to evaluate generative plans. A distribution is useful only if it ranks safe and goal-consistent trajectories above unsafe or irrelevant ones.

— Planning-Oriented End-to-End Autonomous Driving: Architectures, Evaluation, and Emerging Paradigms  (2608.20111 - Guan et al., 20 Aug 2026) in Section 4.4, Uncertainty and Generative Planning

The candidate-oracle regret of 0.0125 shows that better choices exist inside the 32-proposal set, while the representative linear calibration reduces this regret by only 1.06 × 10−4. This identifies unresolved selection headroom; it does not prove a support bottleneck because no external comparison space was retained.

— From Proxy Learning to Driving Decisions: A Transfer-Based Framework for Evaluating Future-Aware Autonomous Driving Planners  (2609.02688 - Wu, 2 Sep 2026) in Section 4.5, “Evaluation-Rule Ablation and Conclusion Reversals”; Section 5.3, “Limitations and Future Work”