Indy Autonomous Challenge Overview

• IAC is an international autonomous racing competition that tests full-scale, high-speed vehicles on iconic circuits like IMS and LVMS.
• The competition employs standardized racecar platforms integrated with advanced sensor suites and modular autonomy stacks for perception, localization, planning, and control.
• Empirical benchmarks demonstrate sub-meter tracking, robust recovery under GNSS outages, and speeds up to 260 km/h, driving innovations for broader autonomous driving applications.

The Indy Autonomous Challenge (IAC) is an international research competition advancing the state of the art in autonomous racing through the deployment of full-scale, high-speed autonomous vehicles on iconic motorsport circuits such as the Indianapolis Motor Speedway (IMS) and Las Vegas Motor Speedway (LVMS). It serves as a rigorous testbed for autonomy stacks operating at the handling limits, fostering developments in multi-modal perception, localization, dynamic planning, robust control, and real-time safety across diverse tracks and complex race formats—including solo time trial and head-to-head racing. The IAC provides a reproducible environment for benchmarking autonomous driving algorithms and hardware, promoting open scientific exchange and catalyzing innovations transferrable to broader autonomous vehicle domains (Ali et al., 23 Sep 2025, Jardali et al., 7 Dec 2025, Jung et al., 2023, Wischnewski et al., 2022, Saba et al., 27 Aug 2024, Rampuria et al., 12 Aug 2024).

1. Competition Structure and Vehicle Platform

The IAC race format has evolved from single-vehicle time trials to multi-agent head-to-head events. Races are typically held on the Dallara AV-21—a standardized racecar platform with a drive-by-wire (DBW) system and instrumentation for high-speed autonomous operation. Teams receive identical hardware suites comprising:

• Compute units: dSPACE AutoBox or equivalent (Intel Xeon CPU, NVIDIA GPUs, high-capacity SSDs).
• Sensor suites: Multi-GNSS RTK/INS (NovAtel, VectorNav), multiple 360° LiDARs (Luminar Iris/Hydra), radar (Continental ARS540), multi-camera arrays.
• Actuation: Schaeffler Paravan SpaceDrive II DBW, Raptor ECU.
• Networking: CAN buses for DBW interface, Gigabit Ethernet/VLAN for sensor domains (Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025, Saba et al., 27 Aug 2024).

Competitions demand precise integration of hardware and software, with practice runs often limited (<15 h real track time), creating a unique constraint on system development and validation (Ali et al., 23 Sep 2025).

2. System Architecture and Real-Time Integration

Most IAC entries deploy a modular, layered architecture based on ROS 2, with rigid node/topic separation and explicit health monitoring. A typical autonomy stack includes:

1. Localization & State Estimation: ESKF or iSAM2 factor graph, fusing GNSS/INS, wheel odometry, LiDAR-odometry.
2. Perception: Sensor fusion of LiDAR, radar, and multi-camera detections for cone/track mapping and opponent recognition.
3. Tracking: EKF/UKF-based multi-object tracking leveraging IMQ or Generalized-Bayes weighting during sensor outages.
4. Planning: Combined global optimization (minimum-curvature raceline) and online local planners (quintic splines, MPC, MPPI).
5. Control: Decoupled or unified feedback (PID, LQR, MPC, Pure Pursuit), tuning for high speeds/handling limits.
6. Safety/Monitor: Watchdogs, real-time state machine, emergency stop routines on critical failure (sensor dropout, localization fault).

Nodes are configured using YAML parameters, executables launched via ros2 launch. Most modules operate at fixed rates: state estimation (100–125 Hz), planning/control (50–100 Hz), perception pipelines callback-driven (20–30 Hz), with health codes/heartbeats propagated for distributed system liveness (Ali et al., 23 Sep 2025, Jardali et al., 7 Dec 2025, Jung et al., 2023, Saba et al., 27 Aug 2024).

3. Multi-Modal Perception and Sensor Fusion

Autonomous vehicles in the IAC rely on multi-sensor suites for detection and classification:

• LiDAR: DBSCAN or learned PointPillars-based clustering (ε ≈ 0.5 m, min_samples = 5), with 360° coverage, cone discrimination up to 35–150 m. Persistence over multiple frames reduces false positives (Abdo et al., 14 Nov 2025, Alvarez et al., 2022, Jardali et al., 7 Dec 2025).
• Camera: YOLOv5/v8 networks trained on domain-specific cone/opponent datasets (FS-OCO, hand-labeled), multi-class detection (e.g., yellow, blue, orange, unknown), [email protected] up to 96%, inference latency <15 ms on dedicated GPUs (Abdo et al., 14 Nov 2025, Rampuria et al., 12 Aug 2024, Alvarez et al., 2022).
• Radar: 4D point cloud returns for static/dynamic object separation, opponent detection up to 200 m. Clustering and Unscented/Extended Kalman Filter for track association (Jardali et al., 7 Dec 2025, Jung et al., 2023).
• Sensor fusion: LiDAR clusters are projected into camera frames using calibrated extrinsics/intrinsics and associated by weighted (distance/confidence) metrics; radar/LiDAR/camera fusion for robust multi-agent tracking. Fusion weights adapt based on environmental confidence, photometric conditions, and expected sensor performance (Abdo et al., 14 Nov 2025, Jardali et al., 7 Dec 2025, Saba et al., 27 Aug 2024).

Perception pipelines demonstrate recall/precision >95% at target detection ranges in controlled conditions, with resilience to lighting/visibility changes (Abdo et al., 14 Nov 2025).

4. Localization, Mapping, and State Estimation

IAC stacks universally adopt probabilistic state estimation integrating exteroceptive and proprioceptive sensors:

• Error-State Kalman Filter (ESKF): Incorporates GNSS position/velocity, IMU angular rates, wheel odometry. Update rate ≥100 Hz, generalized-Bayes/inverse multi-quadratic weighting for outlier mitigation and seamless handover during RTK outages (Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025).
• iSAM2 factor graph SLAM: Incorporates LiDAR scans registered against prior RTK maps, GNSS, and IMU; Bayes-tree optimization for full SLAM–odometry; localization drift <0.5 m during multi-second GNSS loss (Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025).
• Localization performance: Typical RMS lateral error ≲0.15–0.30 m at speeds >200 km/h; heading jerk and position jerk decrease substantially across stack iterations (Figure 1, (Jardali et al., 7 Dec 2025)).
• SLAM front-end: Delaunay triangulation on colored cone maps; reward-based greedy graph search for midpoint extraction, segment angle, and track width scoring (Equation track_width). Spline fitting/velocity profiling via forward-backward sweep, full corridor constraints (Demeter et al., 25 Apr 2025, Alvarez et al., 2022, Jardali et al., 7 Dec 2025).

SLAM and data association are critical for real-time mapping and cone localization. Hybrid FastSLAM 2.0 and Graph SLAM are commonly employed for global consistency and low-latency pose estimation (Demeter et al., 25 Apr 2025, Alvarez et al., 2022).

5. Planning, Trajectory Optimization, and Control Strategies

Two-phase planning modules are universal:

• Offline global planners: Extract raceline via minimum-curvature optimization, solved via quadratic programming; velocity profiling using extended GG/GGS (Grip–G-G) diagrams that account for aerodynamics, tire/friction limits, and mass (Alvarez et al., 2022, Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025).
• Online local planners: Receding-horizon quintic spline extraction, Pure Pursuit on segment midlines, model predictive control (MPC)—linear and nonlinear, including MPPI (Model Predictive Path Integral).
• Behavior logic: Finite-state action selection—maintain raceline, trail/follow opponent, lane change/safe merge (minimum jerk trajectories), attack/overtake. ACC-based curvilinear gap regulation (Equation for $v'_{des}$), safety buffers for collision checks, tire friction constraint enforcement (Jardali et al., 7 Dec 2025, Jung et al., 2023, Saba et al., 27 Aug 2024).

Control stacks encapsulate:

• Lateral control: Pure Pursuit (lookahead $L_d$), Stanley, explicit LQR (state-feedback, velocity-dependent gain scheduling), and full MPC (contouring control). State vector typically $x=[e_y,\dot{e}_y,e_\psi,\dot{e}_\psi]^T$.
• Longitudinal control: PID or P/PI control on velocity error, feedforward compensation, slip-ratio PID for ABS regime, direct torque mapping for DBW interface.
• Stability: Both Pure Pursuit and Stanley are globally asymptotically stable under standard conditions; convex blending and yaw damping terms are often added for high-speed operation (Demeter et al., 25 Apr 2025).
• Failure modes: Safety-critical logic for handling hardware/network failures—brake on localization drop, controlled stop on planner crash, sensor health monitoring, and real-time watchdog counters (Ali et al., 23 Sep 2025, Saba et al., 27 Aug 2024).

Measured control performance includes lateral RMS tracking errors ≲0.7 m, heading error <2°, and acceleration maxima up to 28 m/s² at peak speeds (206–260 km/h) (Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025).

6. Multi-Agent Race, Overtaking Logic, and Safety Supervision

Head-to-head and multi-agent racing introduce new autonomy challenges:

• Opponent detection: Radar+LiDAR fusion, DBSCAN clustering of moving targets, Mahalanobis/gate-based association, unscented Kalman filter for dynamic state estimation (Jardali et al., 7 Dec 2025, Jung et al., 2023).
• Overtaking planners: Rendezvous guidance law (Equation for parallel navigation), minimum jerk trajectory generation in Frenet coordinates, feasibility checks (collision, tire limits, safety margins).
• Safety: System Status Manager (SSM) performs real-time status validation, health checking, fail-safe mode selection (e.g., fallback to LiDAR-only controller), heartbeat/status-code propagation, and external emergency stops (Jung et al., 2023).
• Modular supervision: Distributed state machines watch lower-level ECU/autonomy stack flags, sensor heartbeats, and drive controlled degradations.
• Hardware-in-the-loop: Real-time HIL simulation using duplicate on-board computers, multi-agent vehicle dynamics, and perturbation injection for validation and failover testing (Wischnewski et al., 2022).

At the handling limits, real-time perception, planning, and control must operate within ≤300 ms cycle time to guarantee reaction times and prevent collision. Empirical deployment achieves safe overtakes, repeatable sub-meter errors, and zero race-ending failures in both time-trial and adversarial head-to-head events (Jung et al., 2023, Wischnewski et al., 2022).

7. Empirical Performance, Lessons, and Future Directions

Experimental benchmarks across IAC stacks consistently report:

• Top speeds up to 260 km/h (LVMS), cross-track errors ≲1 m, high-bandwidth feedback, and on-track lap/competition success (Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025).
• Trajectory tracking: Mean lateral error 0.69 m at 60 m/s, lateral acceleration >2 g, heading error ±1–4° (Saba et al., 27 Aug 2024, Jardali et al., 7 Dec 2025).
• Robustness: Detection recall >95%, localization error <0.5 m under GNSS outages, rapid recovery using IMQ weighting or SLAM backup (Jardali et al., 7 Dec 2025).
• Failure modes prompt stack refinement (e.g., watchdog improvements after middleware-induced DBW stop (Ali et al., 23 Sep 2025)).
• Key empirical lessons: Maximum sensor range is as critical as algorithmic complexity; frequent replanning trumps heavy multi-modal optimization in dynamic scenarios; modularity enables rapid deployment and benchmarking (Wischnewski et al., 2022, Saba et al., 27 Aug 2024).
• Limitations: Network dropouts, sensor plate recalibration due to vibration, cold tire-induced spin-out, compute budgeting with multicore CPUs/GPUs.
• Future directions: Integration of tube-MPC for friction limits, game-theoretic or stochastic planning under adversarial uncertainty, real-time learning-based cross-modal fusion, and enhanced multi-agent strategy research (Saba et al., 27 Aug 2024, Jardali et al., 7 Dec 2025).

The IAC has established itself as the de facto testbed for scalable, high-speed autonomous racing, supporting rigorous evaluation of modular stacks, multi-agent competition logic, and robust, real-time safety mechanisms. The competition continues to drive research in perception, planning, control, and system integration under extreme, realistic constraints, with open datasets and reproducible software stacks further enabling empirical advances (Jardali et al., 7 Dec 2025, Ali et al., 23 Sep 2025, Jung et al., 2023, Saba et al., 27 Aug 2024).