Autonomous Race Stack (ARS)

Updated 14 December 2025

Autonomous Race Stack (ARS) is a modular framework combining hardware, software, and algorithms for high-speed, real-time control in racing environments.
It integrates multi-modal sensing, ROS2-based communication, and advanced simulation for rapid deployment and system-level safety.
ARS achieves competitive performance through redundant control systems, robust sensor fusion, and adaptive planning validated in global racing competitions.

The Autonomous Race Stack (ARS) is a modular set of hardware and software architectures, algorithms, and integration methodologies explicitly designed to enable real-time, high-speed, and robust autonomous vehicle control in racing environments. ARS frameworks have been deployed and validated in full-scale head-to-head competitions such as the Indy Autonomous Challenge (IAC), Abu Dhabi Autonomous Racing League (A2RL), Formula Student Driverless, and scalable platforms like F1Tenth; they are characterized by strict real-time requirements, integration of multi-modal sensing (LiDAR, radar, vision), and system-level safety, all under limited on-track validation cycles. ARS encompasses a diverse family of reference implementations, ranging from minimalistic time-trial stacks to highly redundant, supervisory-managed, multi-policy architectures, and unified simulation toolkits. Key contributions from recent works include robust error-state and factor-graph state estimation, minimum-curvature raceline generation, advanced model-based/kinematic/path-following control, safety state machines, and modular rapid deployment in both physical and virtual domains (Ali et al., 23 Sep 2025, Jardali et al., 7 Dec 2025, Bockman et al., 1 Oct 2024).

1. Architectural Foundation and System Hardware

ARS solutions universally employ a modular hardware/software co-design centered around ROS (Robot Operating System) 2 or its analogues for real-time data interchange and decoupled execution. Typical deployments on high-speed race vehicles such as the IAC AV-24 utilize:

Sensors: Multiple GNSS/RTK and IMUs (20–200 Hz), multi-beam LiDARs (10 Hz, 360°), automotive radar (10–20 Hz), multi-camera arrays (10 Hz, 360°, high-res).
Compute: Dedicated real-time controllers (e.g., dSpace AUTERA, Jetson AGX Xavier, Intel Xeon/A5000 GPU combos), SSD storage, high-bandwidth networking (VLANs, CAN, UDP/TCP over SSH).
Actuation: Drive-by-wire integration via CAN, high-speed safety breakouts, and both longitudinal and lateral control with direct throttle/brake/gear actuation.
Interface: Race-control links (vehicle/track flags), base station feedback dashboards, and remote override or emergency stop systems.

Modular software stacks instantiate:

Sensor drivers (C++ nodes), individual sensor time-synchronization, and ROS 2 topics or custom middleware (DDS, ZeroMQ).
Decoupled managers per module minimizing cross-module failures and dropped packets.
Safety/health monitoring at all module boundaries, with supervised state machines for fault containment (Ali et al., 23 Sep 2025, Jardali et al., 7 Dec 2025, Demeter et al., 27 Aug 2024).

2. Perception, Mapping, and State Estimation Algorithms

Core ARS stacks classify environmental perception into pure time-trial (known empty track) versus multi-agent (dynamic obstacles/opponents). In high-speed time-trial configurations, ARS may forego online obstacle detection and semantic segmentation, streaming raw exteroceptive sensor data but assuming a static, pre-mapped track; dynamic racing stacks incorporate multi-sensor fusion for opponent and obstacle detection, using:

LiDAR/Radar Fusion: Clustering (DBSCAN), centroiding, and confidence rating of 3D clusters; SNR and confidence filtering for radar; Kalman or Unscented Kalman motion-state filters; temporal gating for robust track-birth and track-death logic (Ali et al., 23 Sep 2025, Jardali et al., 7 Dec 2025, Saba et al., 27 Aug 2024).
Camera/Deep Learning Perception: YOLOv5/v8 for cone/object detection, monocular/stereo/depth estimation for distance cueing—commonly fused with LiDAR for improved resilience and range. Fusion may be realized by transforming cluster centroids into image space and voting for label consistency (Rampuria et al., 12 Aug 2024, Abdo et al., 14 Nov 2025, Saba et al., 27 Aug 2024).

State Estimation:

EKF/ESKF-based fusion leverages IMU, GNSS/RTK, wheel odometry and, in advanced stacks, robust weighting (inverse multi-quadratic/IMQ) to downweight outliers. When RTK quality degradates, dead-reckoning resets maintain continuity under the small-shift assumption, and factor graph-based SLAM (iSAM2) integrates asynchronous high-rate LiDAR and IMU for drift-free pose and orientation (especially under GNSS denial) (Ali et al., 23 Sep 2025, Jardali et al., 7 Dec 2025).

Offline mapping for time-trial, or GraphSLAM with kNN data association and loop closure for online mapping, underpins precise track-limit extraction and automated reward-based centerline or raceline search (Demeter et al., 25 Apr 2025, Alvarez et al., 2022).

3. Motion Planning and Trajectory Generation

ARS planning decomposes into offline and online phases:

Offline Raceline and Velocity Profile: Boundary extraction via map smoothing (e.g., from Google KML or cone distributions → Delaunay triangulation) and minimum-curvature optimization subject to friction and track constraints. For a given $\kappa(s)$ , the velocity envelope is

$v_{\mathrm{ref}}(s) \leq \sqrt{\frac{\mu g}{|\kappa(s)|}}$

where $\mu$ is friction, $g$ gravity (Ali et al., 23 Sep 2025).

Online Planning: Local planners interpolate to the nearest raceline point (Newton–Raphson), forward-sample along the curve (typically horizon-fixed), enforce race-control-imposed flags, and produce local waypoints in vehicle frame for path tracking (Ali et al., 23 Sep 2025, Jardali et al., 7 Dec 2025). In head-to-head or reactive stacks, local planners generate dynamic overtaking or evasion trajectories (e.g., quintic/quartic polynomials in Frenet or multi-spline blends) that guarantee collision checking, curvature, and boundary feasibility (Jung et al., 2023, Baumann et al., 18 Mar 2024).

Multi-policy Supervisory ARS: Modern implementations instantiate several trajectory generation pipelines in parallel (geometric, learning-based, teleop, RL, etc.), with a decision logic and “clutch” module arbitrating transitions under robust-invariance constraints to ensure safety and minimize control discontinuity (Demeter et al., 27 Aug 2024).

4. Control Algorithms: Longitudinal and Lateral Strategies

Longitudinal Control: PID or PI cascades, sometimes with feed-forward drag terms, track $v_{\mathrm{ref}}$ ; anti-lock brake systems (ABS) maintain optimal slip ratios with logic such as

$\lambda = \frac{V_{\mathrm{veh}} - \omega R}{V_{\mathrm{veh}}}$

for slip $\lambda$ , wheel speed $\omega$ , and radius $R$ (Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025). Velocity profiles are imposed as hard constraints, modulated for flags and safety events.

Lateral Control:

Pure Pursuit: Computes steering commands for arc tracking of look-ahead waypoints:

$\delta = \arctan\left(\frac{2L \sin \alpha}{L_d}\right)$

with $L$ wheelbase, $\alpha$ heading error, $L_d$ look-ahead distance.

LQR/MPC: Linear Quadratic Regulator (LQR) on linearized dynamic bicycle model states, or Model Predictive Control (MPC) in lateral error coordinates; cost functions

$J = \int_0^\infty x^T Q x + u^T R u\,dt$

with online scheduling of $Q, R$ across velocity bands, or tube/hybrid MPC constraints (Saba et al., 27 Aug 2024, Jung et al., 2023).

Advanced/Planned Integration: Some stacks deploy GPU-parallel Model Predictive Path Integral Control (MPPI) over cost functionals in multi-body dynamic state, or blended controllers (e.g., adaptive Stanley/Pure Pursuit weighting) for optimal tracking (Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025, Baumann et al., 18 Mar 2024).

5. Integration, Validation, and System-Level Safety

System Integration: Virtual-in-the-loop simulation using Unity-based environments (AWSIM), distributed managers per sensor/module (non-monolithic launch), and full-stack calibration including track banking priors and multi-IMU extrinsic routines (Yu et al.). Parameter scheduling proceeds from low-speed open-loop to full-speed closed-track runs, with typically only a few parameter adjustments per track session (Ali et al., 23 Sep 2025).

Rapid Deployment: Successful stacks (e.g., 206 km/h top speed at IMS within 11 h or 325 km on-track, as little as two weeks to port scaled architectures to new vehicles) demonstrate a minimal integration regime focused on leveraging robust standard modules, modular pipeline managers, and early comprehensive simulation (Ali et al., 23 Sep 2025, Demeter et al., 27 Aug 2024, Jardali et al., 7 Dec 2025).

Safety Architecture: Multiplexed heartbeat timeouts at every module, direct cross-track error fences (e.g., $|e_{ct}|>3.5$ m triggers stop), sensor health gating, rolling counter validation on CAN, and remote override/kill capability at base station (Ali et al., 23 Sep 2025). Supervisory-managed ARS with formal state machines enable rapid safe fallback to backup pipelines or control regimes (Demeter et al., 27 Aug 2024).

6. Quantitative Performance, Datasets, and Benchmarks

Benchmark ARS stacks have achieved:

Metric	Value/Range	Reference
Top Speed	205–260 km/h (full-scale)	(Ali et al., 23 Sep 2025, Jardali et al., 7 Dec 2025)
Cross-track error	$[-1.4, +0.8]$ m (@200km/h)	(Ali et al., 23 Sep 2025)
Heading error	$[-2.0^\circ, +2.25^\circ]$	(Ali et al., 23 Sep 2025)
Lateral acceleration peak	$18$–$28$ m/s² ( $\sim2.4g$ )	(Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025)
Perception latency	$<40$ ms (LiDAR/vision fusion)	(Rampuria et al., 12 Aug 2024, Abdo et al., 14 Nov 2025)
Absolute localization error	$<0.15$ m after $30$ cones (SLAM)	(Rampuria et al., 12 Aug 2024)
Lap/track average speed	$\sim200$ km/h	(Ali et al., 23 Sep 2025)
Pipe-switch latency, robustness	Smooth clutch, no single-point failure	(Demeter et al., 27 Aug 2024)
Dataset availability	Open multi-sensor racing sets	(Jardali et al., 7 Dec 2025)

Multiple stacks provide open-access, high-speed, multi-sensor datasets from IMS, LVMS, and LVRC, including aligned LiDAR, GNSS/IMU, radar, camera, raceline trajectories, and opponent ground truth (Jardali et al., 7 Dec 2025).

7. Lessons Learned and Future Directions

Minimalism Outperforms Complexity Under Resource Constraints: Lean stacks with robust, simple modules (standalone drivers, PID/LQR, offline raceline) achieve competitive results with minimized integration risk (Ali et al., 23 Sep 2025).
Supervisory Redundancy Is Essential for Robustness: Parallel pipelines with formal supervisors mitigate single-point-of-failure risks, facilitate experimental agility, and enable on-the-fly fallback (Demeter et al., 27 Aug 2024).
Planning/Control Coupling Requires Accurate Modeling: Decoupled PID/MPC induces instability at dynamics limits (e.g., ABS-induced spin), motivating coupled stochastic controllers (e.g., MPPI) (Jardali et al., 7 Dec 2025).
Real-World Constraints Dominate Stack Design: Limited on-track access, variable hardware environments, and in situ failures drive iterative tuning and favor simulation-backed, modular stacks (Ali et al., 23 Sep 2025, Jardali et al., 7 Dec 2025).
Sensor Fusion/Calibration Remain Bottlenecks: Cross-modal alignment of LiDAR, camera, GNSS, IMU, and wheel odom is a principal challenge, especially with dynamic reconfiguration and under varying environmental conditions (Jardali et al., 7 Dec 2025, Saba et al., 27 Aug 2024).
Path Forward: Ongoing work progresses toward online self-calibration, real-time multi-agent planning under robust vehicle-tire model uncertainty, formal safety proofs for supervisory logic, and further scalable reduction of entry barriers via open, reproducible simulation stacks (Jardali et al., 7 Dec 2025, Bockman et al., 1 Oct 2024, Abdo et al., 14 Nov 2025).