Results of the 2024 CommonRoad Motion Planning Competition for Autonomous Vehicles

Yanliang Huang, Xia Yan, Peiran Yin, Zhenduo Zhang, Zeyan Shao, Youran Wang, Haoliang Huang, Matthias AlthoffarXiv 2025 · DOI · arXiv

Introduction

The 2024 CommonRoad Motion Planning Competition highlights a critical trade-off in autonomous driving: the balance between "solving" a wide variety of complex scenarios (coverage) versus the "quality" of the driving itself (efficiency, safety, and comfort). The winning entry from the Technical University of Munich (TUM) demonstrated significant progress in scenario coverage, successfully navigating 31.1% more scenarios than the previous year's champion by utilizing a hybrid planning approach.

The Benchmarking Challenge

Despite rapid progress in autonomous vehicle (AV) research, the field has historically lacked standardized, reproducible benchmarks. Many motion planning algorithms are evaluated on private datasets or within specific simulators, complicating fair comparisons. The CommonRoad competition addresses this by providing an open-source framework where planners are executed on identical hardware against a set of realistic, high-stakes scenarios. These scenarios include highway and urban environments with diverse participants like buses and bicycles, requiring planners to balance efficiency, safety, comfort, and strict traffic rule compliance.

Methodology: Hybrid Planning and Game Theory

The 2024 winning planner (TUM-2024) employs a dual-module architecture designed to handle different driving contexts:

  1. Modified Reactive Planner: For general road segments, the system uses a sampling-based planner formulated in the Frenét frame. It generates candidate trajectories using quintic curves, evaluated based on kinematic feasibility and a cost function. It integrates Signal Temporal Logic (STL) robustness monitors to ensure compliance with traffic rules, such as maintaining safe following distances.
  2. Level-$k$ Game Theory Planner: To manage high-stakes interactions at unsignalized intersections, the team implemented a game-theoretic model. This approach assumes other drivers have different "reasoning depths" (Level-0 to Level-2). The ego vehicle maintains and updates a "belief distribution" about other agents' reasoning levels in real-time, enabling more human-like, interaction-aware decision-making.
Progression of a traffic scenario over time in the CommonRoad benchmark suite.

Evaluation and Results

The evaluation utilized the CommonRoad benchmark suite, consisting of over 500 scenarios derived from real-world traffic recordings in German cities like Aachen, Cologne, and Dresden. Planners were tested using a nonlinear vehicle model on interactive benchmarks using the SUMO traffic simulator.

Coverage vs. Quality Trade-off

A key insight from the 2024 results is the divergence between the TUM-2024 winner and the 2023 winner (Stony Brook University - SB).

  • Scenario Coverage: TUM-2024 focused on breadth, solving 112 scenarios that SB-2023 could not. This represents a 31.1% improvement in coverage, largely due to its computational efficiency within the competition's 6-hour limit.
  • Trajectory Quality: Conversely, SB-2023 focused on depth. On the 131 scenarios solved by both planners, SB-2023 achieved approximately 40% lower average costs (TR1 metric), indicating superior comfort and rule compliance.

Conclusion and Future Directions

The success of the Level-$k$ model suggests that explicit interaction modeling is crucial for complex environments like intersections. However, the TUM-2024 planner's reliance on simplified kinematic predictions led to overly conservative behavior. Future work aims to bridge the gap between TUM’s robust coverage and SB’s high-fidelity trajectory optimization, potentially by integrating deep-learning-based prediction models to reduce conservatism and improve overall driving quality.

Summary updated

BibTeX
@misc{huang2025results,
  author = {Yanliang Huang and Xia Yan and Peiran Yin and Zhenduo Zhang and Zeyan Shao and Youran Wang and Haoliang Huang and Matthias Althoff},
  title = {{Results of the 2024 CommonRoad Motion Planning Competition for Autonomous Vehicles}},
  year = {2025},
  eprint = {2512.19564},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.19564},
}
Figure