AROLA: A Modular Layered Architecture for Scaled Autonomous Racing

Technical Summary: AROLA – A Modular Layered Architecture for Scaled Autonomous Racing

Problem Statement
The field of scaled autonomous racing, particularly on platforms like RoboRacer (formerly F1Tenth), has seen rapid growth. However, the software stacks developed for these systems often suffer from fragmentation, monolithic designs, and tight coupling to specific hardware or competition rules. This lack of modularity hinders reusability, complicates fair benchmarking across research groups, and makes systematic evaluation difficult. While end-to-end learning approaches offer alternatives, they often sacrifice interpretability and maintainability. There is a critical need for an open, layered framework that standardizes interfaces, allows for interchangeable components, and supports reproducible evaluation without relying on custom message definitions.

Methodology
The authors propose AROLA (Autonomous Racing Open Layered Architecture), a modular, open-layered software architecture designed to decompose the autonomous driving pipeline into eight distinct, interchangeable layers connected via standardized ROS 2 interfaces. The architecture follows a top-to-bottom data flow with conservative cross-layer interactions to preserve modularity:

1. Sensing: Acquisition of raw environmental data via onboard sensors.
2. Preprocessing: Filtering, cleaning, and fusing raw data streams to reduce noise and computational overhead.
3. Perception: Assigning semantic meaning to data (e.g., identifying track boundaries and obstacles).
4. Localization and Mapping: A coupled process (SLAM) for determining vehicle pose and constructing maps.
5. Planning: Generating feasible trajectories based on perception and map data.
6. Behavior: Governing decision-making logic (e.g., via state machines or behavior trees) to determine what the vehicle should do.
7. Control: Executing motion plans and compensating for actuation imperfections.
8. Actuation: Physical execution of commands on hardware.

To support systematic evaluation, the authors introduce Race Monitor, a lightweight, ROS 2-based monitoring framework. Race Monitor logs real-time telemetry (lap times, trajectory quality, computational load) and generates standardized post-race analyses using the evo package. It publishes data to standard ROS 2 topics, allowing controllers to subscribe without introducing coupling, and supports various trajectory formats (CSV, TUM, KITTI).

Key Contributions
The paper presents three primary contributions:

1. AROLA: A modular, open-layered software architecture that standardizes interfaces (using standard ROS 2 message types like LaserScan, Odometry, OccupancyGrid, and AckermannDriveStamped) to facilitate component interchangeability and rapid module replacement.
2. Race Monitor: An integrated real-time monitoring and benchmarking suite that enables objective performance evaluation, trajectory error visualization, and consistent comparison of control strategies.
3. Validation: Comprehensive experimental validation of the framework in both simulated environments and on physical hardware using the RoboRacer platform.

Experimental Results
The framework was validated during the preparation for the 2025 RoboRacer IV25 competition and in post-competition experiments:

• Controller Comparison: Three controllers (Gap Follower, Pure Pursuit, and MPC) were evaluated. Pure Pursuit achieved the fastest lap times (10.10 s best lap) with moderate computational load (28% CPU) and control latency (18 ms), outperforming MPC in speed despite MPC's lower jerk. Gap Follower was the slowest but most reactive.
• Competition Performance: The finalized AROLA configuration secured third place in the RoboRacer IV25 competition with a lap time of 10.1 s.
• LQR Evaluation: A post-competition evaluation of an LQR controller on a "Berlin Map" demonstrated the system's ability to log detailed metrics. The LQR controller completed seven laps with a best time of 19.53 s and a high consistency score (0.99), with Race Monitor successfully identifying trajectory deviations and scaling misalignments.

Significance and Claims
The authors claim that AROLA and Race Monitor demonstrate that modularity, transparent interfaces, and systematic evaluation can accelerate development and improve reproducibility in scaled autonomous racing. By reorganizing fragmented designs into standardized layers, the architecture lowers the barrier to entry while preserving the expressive power needed for advanced racing. The framework allows teams to swap components (e.g., changing controllers or perception modules) without reworking the entire stack and enables unbiased comparison of strategies under identical conditions. The paper concludes that this approach provides a practical foundation for quick iteration, fair performance evaluation, and dependable scaling in autonomous racing research.

Limitations and Future Work
The authors acknowledge that the current evaluation is limited in scope, focusing primarily on a small set of controllers rather than the full range of potential modules. The system remains susceptible to error propagation across pipeline stages, and validation is currently centered on a specific hardware platform. Future work aims to improve tooling for coordinate-frame alignment, extend validation to diverse hardware and simulators, develop a watchdog system for failure prevention, and create a web-based interface for Race Monitor. Additionally, the authors plan to explore integrating learning-based or partial end-to-end components within the modular framework while maintaining standardized interfaces.