CommonRoad Motion Planning Competition

• The CommonRoad competition is an annual event that benchmarks autonomous vehicle motion planners with standardized, reproducible scenarios and strict safety metrics.
• It employs a two-phase format featuring open didactic and closed evaluation scenarios, integrating real-time SUMO simulations for dynamic multi-agent interactions.
• Benchmark metrics including trajectory cost, safety, comfort, and rule compliance enable rigorous and quantitative comparisons across diverse algorithmic strategies.

The CommonRoad Motion Planning Competition is an annual open, reproducible benchmarking event for evaluating autonomous vehicle motion planning algorithms under realistic, diverse, and safety-critical scenarios. Conducted using the CommonRoad benchmark suite, the competition systematically compares algorithms along standardized metrics encompassing efficiency, safety, comfort, and rule compliance, establishing a rigorous platform for quantitative assessment and development in the autonomous driving community (Kochdumper et al., 2024, Huang et al., 22 Dec 2025).

1. Competition Format and Benchmarking Workflow

The competition operates in a two-phase format:

• Phase I ("Didactic"): Open access to known, public CommonRoad scenarios—both non-interactive and interactive—allowing teams to familiarize themselves with the benchmark structure and evaluation process.
• Phase II (Evaluation): Registered teams submit Docker images implementing their planners against a closed set of non-public, interactive scenarios. Submissions are executed on standardized hardware (maximum two CPU cores, six-hour wall-clock limit) within the CommonRoad evaluation server farm.

Interactive scenarios employ realistic road networks (lanelets), and other agents (cars, buses, bicycles, and sometimes pedestrians) are simulated using the SUMO traffic simulator, allowing them to react in real-time to the ego vehicle's maneuvers. The strict separation and reproducibility of scenarios and solver environments facilitate unbiased comparative analysis across institutions and years (Kochdumper et al., 2024, Huang et al., 22 Dec 2025).

2. CommonRoad Benchmark Suite and Scenario Design

The CommonRoad suite incorporates over 500 scenarios spanning both highway (multi-lane, high-speed, ramps) and urban (intersections, complex layouts, varied participants) environments. Maps use real road networks, with explicit annotation of drivable lanelets, sidewalks, bicycle-only lanes, explicit road compliance regions, and forbidden zones.

Obstacles and traffic agents in these scenarios include:

• Static: lane boundaries, edges, and crosswalks.
• Dynamic: passenger cars, buses, trucks, bicycles, and, in select urban settings, virtual pedestrians using linear or kinematic single-track vehicle models derived from real-world vehicles.
• In interactive scenarios, other agents are controlled by SUMO and respond dynamically to the ego trajectory, simulating multi-agent interactivity of realistic traffic (Kochdumper et al., 2024, Huang et al., 22 Dec 2025).

3. Evaluation Metrics and Cost Function Design

Solutions are first screened for feasibility—collision avoidance (no ego-obstacle overlap), kinematic feasibility (respecting the nonlinear single-track vehicle model), and road compliance (remaining inside drivable lanelets).

For feasible, goal-reaching trajectories, the scalar cost function $J_\text{ego}$ (also termed TR1) is defined as:

$$J_\text{ego} = w_1 J^\text{lon\_J} + w_2 J_{SR} + w_3 J_{D} + w_4 J_{LC}$$

with weights $w=[0.01, 22, 8, 5]$ (Huang et al., 22 Dec 2025), where:

• Longitudinal jerk penalty ($J^\text{lon\_J}$):

$$J^\text{lon\_J} = \int_{t_0}^{t_f} \big( \frac{d^3s}{dt^3} \big)^2 dt$$

• Steering-rate penalty ($J_{SR}$):

$J_{SR} = \int_{t_0}^{t_f} v_\delta(t)^2 dt$

• Proximity to obstacles ($J_{D}$):

$$J_D = \int_{t_0}^{t_f} \max_{i} \exp(-w_\text{dist} d_i(t)) dt, \quad w_\text{dist} = 0.2$$

with $d_i$ the Euclidean gap to obstacle $i$.

• Lane-centering offset ($J_{LC}$):

$J_{LC} = \int_{t_0}^{t_f} d(t)^2 dt$

where $d(t)$ is the lateral distance to the lane center.

Additional dimensions include:

• Efficiency: time to reach the goal or number of benchmarks solved.
• Safety: zero collisions, distribution of time-to-collision (TTC) close calls.
• Comfort: low jerk and steering rates.
• Compliance: monitored via signal-temporal logic (STL) for rules such as safe following distance (R_G1) and traffic flow (R_G4); violations increase cost or lead to disqualification (Huang et al., 22 Dec 2025).

4. Algorithmic Strategies of Top-Performing Planners

Several diverse paradigms have demonstrated competitive performance:

Planner (Year/Affiliation)	Principal Approach	Notable Features
SB-2023 (Stony Brook)	Reachability-based corridor selection + OCP	Modular separation, rigorous constraint enforcement
FRENETIX (TUM-2023)	Sampling-based planning in Frenet coordinates	Highly modular, optional ML-based components
TUM-2024 (Yanliang Huang et al.)	Reactive sampling + level-$k$ game-theory for intersections	Adaptive replanning, explicit STL rule monitoring
Trapezoidal Prism Corridor (Deolasee et al., 2022)	Spatio-temporal convex corridor + continuous-time Bézier optimization	Theoretically guaranteed safety, high success rate

• SB-2023: High-level reachability analysis constructs feasible driving corridors by double integrator propagation, then enforces traffic rules and driving constraints; a nonlinear OCP optimizes the trajectory within the corridor, attaining low comfort cost and conservative safety.
• FRENETIX (TUM-2023): Samples quintic/cubic polynomial terminal states in Frenet frame; each candidate undergoes kinematic and rule checks before scoring. Machine learning add-ons for prediction or risk fields can be enabled.
• TUM-2024: Employs a 3-second horizon, replanned every 0.3 s. Intersections invoke level-$k$ dynamic games (14 discrete maneuvers, $k=0,1,2$), using reward functions penalizing collision, off-road, and lane deviation, with belief updates based on $L_1$-norm differences. STL monitors enforce rules in real time (Huang et al., 22 Dec 2025).
• Trapezoidal Prism/Bézier (Editor’s term, (Deolasee et al., 2022)): Generalizes axis-aligned (cuboidal) corridors to linearly time-varying (trapezoidal) corridors in $s$ (longitudinal) and $l$ (lateral), enabling piecewise Bézier curve optimization. Continuous-time collision avoidance is enforced via convex-hull constraints on Bézier control points within the prism, empowering the approach to handle both moving and static obstacles on arbitrary maps.

5. Quantitative Results and Comparative Analysis

Recent competitions (2023–2024) illustrate critical trade-offs:

• Coverage (solved benchmarks): TUM-2024 achieved ≈31% higher scenario coverage versus SB-2023 (TUM-2024: 112 solved only by TUM, SB-2023: 55 only by SB, both: 131 out of 360).
• Trajectory Quality: On common solved benchmarks, SB-2023 consistently achieved a $\approx40\%$ lower mean cost (TR1), lower variance, and better worst-case trajectory cost.
• Mean Comfort and Safety: SB-2023 produced lower mean jerk/steering cost and fewer sub-second TTC events, indicating a more conservative planning style.
• Efficiency and Scalability: TUM-2024’s sampling/reactive paradigm solved more benchmarks within the fixed compute budget, despite yielding higher per-trajectory cost.
• Real-time Constraints: Solvers requiring over one second per replanning iteration accumulated fewer solved scenarios due to the strict wall-clock limit (Kochdumper et al., 2024, Huang et al., 22 Dec 2025).

6. Methodological Advances from Benchmarking

Key algorithmic and architectural advances include:

• Hybridization: Combining efficient corridor-based reachability analysis with real-time sampling for local minima escape to exploit both scalability and solution smoothness (Kochdumper et al., 2024).
• Rule Monitoring: STL-based tools for continuous-time rule robustness monitoring, enabling implementers to penalize or reject trajectories violating safe distance or traffic flow regulations (Huang et al., 22 Dec 2025).
• Safety Guarantees: Use of control point convex-hull conditions with Bézier curves in trapezoidal corridors facilitates continuous-time, interpretable safety guarantees, and robust handling of non-trivial timing/dynamics constraints (Deolasee et al., 2022).
• Learning Integration: Machine learning modules are deployed for motion prediction, risk field estimation, or adaptive cost weight tuning; reinforcement learning is suggested for online meta-optimization in repeated layouts (Kochdumper et al., 2024).

7. Challenges, Lessons, and Future Directions

Observed challenges and recommendations include:

• Prediction and Uncertainty: The need to integrate advanced outcome prediction models and explicit risk/uncertainty quantification to reduce conservativeness without compromising safety.
• Scenario Diversity: Calls for more non-motorized agents, unprotected left turns, and adversarial (“outlier”) scenarios to better stress algorithm robustness (Kochdumper et al., 2024).
• Formal Rule Compliance: Expansion from the two currently monitored rules (R_G1 and R_G4) to richer sets via formal verification or runtime STL monitoring.
• Real-time Adaptability: Enhanced algorithmic constructs for rapid adaptation under compute constraints in densely interactive scenes.
• Scalability versus Per-Trajectory Optimality: The benchmarking results highlight a persistent trade-off—coverage and speed (sampling-based, e.g., TUM-2024) versus solution optimality and conservatism (optimization-based, e.g., SB-2023). Conceivable future work involves combining learning-based risk fields, formal rule monitors, and GPU-accelerated optimization for adaptive, scalable, and certifiable planning (Huang et al., 22 Dec 2025, Deolasee et al., 2022).

The CommonRoad Motion Planning Competition establishes a reference point for evaluating progress in autonomous vehicle planning, highlighting the value of standardization, open data, and reproducibility for robustly comparing diverse approaches and motivating innovative solutions to the remaining open challenges in autonomous driving (Kochdumper et al., 2024, Huang et al., 22 Dec 2025, Deolasee et al., 2022).