A Simulation-First ROS 2 Platform for Autonomous Formula Student Vehicles: Mission Control, EKF State Estimation and Statistically Validated Testing
- 1K.J. Somaiya School of Engineering, Somaiya Vidyavihar University
Abstract
Formula Student (FS) Driverless teams must validate a perception-to-control autonomy stack long before the car's sensors and compute exist, yet published systems generally assume the hardware is already on the vehicle. This paper presents a simulation-first platform on ROS~2 Humble and the eufs\_sim2 simulator in which closed-loop controllers for all four FS Driverless events are developed, tested and statistically characterised before hardware is procured. The autonomy package depends on the simulator only through a message-definition package, so the same code is designed to run on a Jetson-based vehicle unchanged. We implement proportional speed control, pure-pursuit steering and perception-gated lap counting, integrate an Extended Kalman Filter (EKF) that fuses IMU and wheel odometry, and build a software-in-the-loop (SIL) harness with an injectable sensor-error model. A Monte-Carlo method with Wilson-score intervals shows Skidpad success of 100\% (95\% CI 67.6--100\%, n=8) up to 8∘/√(min) of odometry yaw drift and 87.5\% (CI 52.9--97.8\%) at 10∘/√(min), where the worst episode reaches 3.84~m although the median stays at 0.44~m. The SIL model predicted the full-simulator mean tracking error to within 0.01~m (0.32 vs.\ 0.33~m), and the EKF drifted 2.4\% over an 83~m run. We also document four defects in the upstream simulator, two of which fail silently. All results are from simulation; camera and LiDAR fusion are not part of the evaluated pipeline.
Related
- Cost-effective nanorough black nickel-based SERS substrates with statistically validated spatial uniformity
- ASTRA: An Autonomous Framework for Physics-Validated Analysis and Hypothesis Generation in Astrophysics
- Agent-E2MD: Autonomous Translation of Interatomic Potential Equations into Physically Validated Pair Styles for Molecular Dynamics in LAMMPS
- MolecureAI: an agentic AI platform for autonomous drug discovery
- Bridging the gap: a genetically validated avian chorioallantoic membrane platform for investigation of spontaneous circulating tumor cells
BibTeX
@misc{satish2026simulationfirst,
title = {{A Simulation-First ROS 2 Platform for Autonomous Formula Student Vehicles: Mission Control, EKF State Estimation and Statistically Validated Testing}},
author = {Rishi Satish and Harshvardhan Singh and Ayon Majumdar and Makarand Govind Kulkarni},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23493501},
url = {https://zenodo.org/records/23493501},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - A Simulation-First ROS 2 Platform for Autonomous Formula Student Vehicles: Mission Control, EKF State Estimation and Statistically Validated Testing
AU - Rishi Satish
AU - Harshvardhan Singh
AU - Ayon Majumdar
AU - Makarand Govind Kulkarni
PY - 2026
DA - 2026/10/20
PB - Zenodo
DO - 10.5281/zenodo.23493501
UR - https://zenodo.org/records/23493501
AB - Formula Student (FS) Driverless teams must validate a perception-to-control autonomy stack long before the car's sensors and compute exist, yet published systems generally assume the hardware is already on the vehicle. This paper presents a simulation-first platform on ROS~2 Humble and the eufs\_sim2 simulator in which closed-loop controllers for all four FS Driverless events are developed, tested and statistically characterised before hardware is procured. The autonomy package depends on the simulator only through a message-definition package, so the same code is designed to run on a Jetson-based vehicle unchanged. We implement proportional speed control, pure-pursuit steering and perception-gated lap counting, integrate an Extended Kalman Filter (EKF) that fuses IMU and wheel odometry, and build a software-in-the-loop (SIL) harness with an injectable sensor-error model. A Monte-Carlo method with Wilson-score intervals shows Skidpad success of 100\% (95\% CI 67.6--100\%, n=8) up to 8^∘/√(min) of odometry yaw drift and 87.5\% (CI 52.9--97.8\%) at 10^∘/√(min), where the worst episode reaches 3.84~m although the median stays at 0.44~m. The SIL model predicted the full-simulator mean tracking error to within 0.01~m (0.32 vs.\ 0.33~m), and the EKF drifted 2.4\% over an 83~m run. We also document four defects in the upstream simulator, two of which fail silently. All results are from simulation; camera and LiDAR fusion are not part of the evaluated pipeline.
ER -