Perfect Prediction or Plenty of Proposals? What Matters Most in Planning for Autonomous Driving

Aron Distelzweig, Faris Janjos, Oliver Scheel, Sirish Reddy Varra, Raghu Rajan, Joschka BoedeckerarXiv 2025 · DOI · arXiv

Abstract

This paper challenges the prevailing assumption in autonomous driving (AD) that more accurate motion predictions for surrounding agents necessarily lead to better planning. Through a series of "perfect prediction" stress tests on state-of-the-art integrated prediction and planning (IPP) models, the authors demonstrate that the current bottleneck is not prediction accuracy, but rather the quality and diversity of the ego-vehicle's own trajectory proposals. To address this "proposal gap," the authors introduce SPDM (Spline-Fitting PDM), a rule-based planner that generates a wide array of diverse, physically feasible trajectories. SPDM achieves state-of-the-art performance on the challenging nuPlan interPlan benchmark, outperforming complex neural planners by roughly 50% in interactive scenarios.

Introduction

The industry is shifting from modular pipelines toward Integrated Prediction and Planning (IPP), where systems jointly reason about the ego-vehicle and surrounding agents. The core hypothesis is that high-fidelity predictions of others' behaviors should enable safer, more human-like plans. However, the authors observed a paradox: even when integrated models are provided with ground-truth future information, their performance often stagnates or declines. This suggests that the real challenge lies in how planners generate and select their own options.

The "Perfect Prediction" Paradox

The researchers evaluated three representative IPP models—DTPP, GameFormer, and DIPP—under three conditions: learned predictions, perfect ground-truth predictions, and no predictions.

Figure 1: The "Perfect Prediction" paradox. (a) shows a collision occurring even when the DTPP planner is given perfect ground-truth predictions. (b) shows the same model avoiding a collision using its own learned predictions, highlighting that the planner fails to effectively utilize high-quality future data.

The study found that for many state-of-the-art models, prediction accuracy is not the limiting factor. In several interactive scenarios, providing ground-truth trajectories actually resulted in lower planning scores. This indicates a fundamental inability of current neural planners to translate accurate environmental forecasts into successful maneuvers.

Bridging the "Proposal Gap" with SPDM

The authors identify the "proposal gap" as the true bottleneck: many planners fail to even consider the correct maneuver (e.g., a simple lane change) in complex settings. To solve this, they developed SPDM (Spline-Fitting PDM), a lightweight planner that prioritizes proposal diversity.

SPDM improves upon the standard Predictive Driver Model (PDM) through:

  • Spline-Based Path Generation: Using cubic polynomials to generate diverse paths across all available lanes, moving beyond simple centerline following.
  • Velocity and Offset Diversity: Generating candidates with target velocities ranging from 20% to 100% of the speed limit and various lateral offsets.
  • Multi-Objective Scoring: Selecting the optimal trajectory based on collision avoidance, passenger comfort, and progress.
Figure 4: Qualitative comparison of planning proposals. While learning-based models (a-d) often struggle to propose viable paths in dense traffic, SPDM (e) generates a rich set of candidates that allow for safe navigation.

Key Results and Insights

The evaluation on the nuPlan framework, specifically the interPlan benchmark for "long-tail" interactive scenarios, yielded several key findings:

  • SOTA Performance: SPDM set new state-of-the-art results on interPlan, significantly outperforming complex models like DiffusionPlanner and GameFormer.
  • 50% Improvement: In interactive scenarios, SPDM improved planning scores by approximately 50% over the original PDM baseline.
  • Rule-Based Robustness: Unlike neural planners that may "hallucinate" unrealistic behaviors in out-of-distribution scenarios, SPDM's physics-based approach ensures feasible and jitter-free trajectories.
  • Efficiency: SPDM remains computationally efficient, scaling linearly with the number of proposals and meeting real-time production requirements.

Conclusion and Future Work

This work demonstrates that for current AD systems, the diversity of trajectory proposals matters more than the precision of agent predictions. While SPDM proves the power of a "plenty of proposals" approach, it still relies on hand-tuned scoring. Future research should explore hybrid architectures that use machine learning to refine the selection process from a robust set of rule-based proposals, and investigate how models can be designed to truly leverage high-fidelity predictions in a closed-loop manner.

Summary updated

Abstract

Traditionally, prediction and planning in autonomous driving (AD) have been treated as separate, sequential modules. Recently, there has been a growing shift towards tighter integration of these components, known as Integrated Prediction and Planning (IPP), with the aim of enabling more informed and adaptive decision-making. However, it remains unclear to what extent this integration actually improves planning performance. In this work, we investigate the role of prediction in IPP approaches, drawing on the widely adopted Val14 benchmark, which encompasses more common driving scenarios with relatively low interaction complexity, and the interPlan benchmark, which includes highly interactive and out-of-distribution driving situations. Our analysis reveals that even access to perfect future predictions does not lead to better planning outcomes, indicating that current IPP methods often fail to fully exploit future behavior information. Instead, we focus on high-quality proposal generation, while using predictions primarily for collision checks. We find that many imitation learning-based planners struggle to generate realistic and plausible proposals, performing worse than PDM - a simple lane-following approach. Motivated by this observation, we build on PDM with an enhanced proposal generation method, shifting the emphasis towards producing diverse but realistic and high-quality proposals. This proposal-centric approach significantly outperforms existing methods, especially in out-of-distribution and highly interactive settings, where it sets new state-of-the-art results.
BibTeX
@misc{distelzweig2025perfect,
  author = {Aron Distelzweig and Faris Janjos and Oliver Scheel and Sirish Reddy Varra and Raghu Rajan and Joschka Boedecker},
  title = {{Perfect Prediction or Plenty of Proposals? What Matters Most in Planning for Autonomous Driving}},
  year = {2025},
  eprint = {2510.15505},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.15505},
}
Figure