Development of a full-stack RoboRacer autonomous racing system capable of overtaking in a head-to-head race

The racing environment often serves as a driving factor for the enhancement of current systems in a field. The rising popularity of autonomous vehicles has naturally led to the development of autonomous racing platforms to push the limits of current technologies and algorithms. These algorithms focus on perception, planning, and control in highly dynamic and uncertain environments, requiring real-time decision-making and adaptability. The autonomous racing environment provides an ideal testbed for developing and evaluating these algorithms as not only does it provide a closed environment with limited variables, but also demonstrates the behaviour of the autonomous systems at its limits. The autonomous racing problem can be summarised as for a vehicle to safely navigate a racetrack, avoiding collisions with opponent vehicles, as quickly as possible. This thesis presents the design and implementation of a full-stack autonomous racing system for the RoboRacer platform, capable of detecting, tracking, and overtaking opponent vehicles in multi-vehicle racing scenarios. The system integrates three primary modules: perception, planning, and control, following the classical “See-Think-Act” paradigm. In the perception module, Monte Carlo localisation (MCL) is used to estimate the ego vehicle’s pose by fusing light detection and ranging (LiDAR) and odometry data, while an adaptive breakpoint clustering (ABC) algorithm combined with a particle filter (PF) provides robust opponent detection and tracking. The hierarchical planning module consists of a global planner that generates a minimum-curvature raceline, a behavioural planner based on a finite state machine (FSM) for high-level overtaking decisions, and a local planner using the rapidly-exploring random tree star (RRT*) algorithm to compute collision-free trajectories around dynamic obstacles. The control module computes the necessary control commands to follow the planned trajectory. Three different control strategies are implemented and compared: Pure Pursuit, Stanley, and model predictive control (MPC) for trajectory tracking and speed regulation with Pure Pursuit being the most effective in this context. The system was developed and evaluated in both simulated and physical environments. The perception subsystem achieved reliable opponent state estimation with a position RMSE of 0.15 m and a speed RMSE of 0.31 m/s on physical vehicles on a racetrack of size 9x5 meters. The planning and control modules enabled successful overtaking manoeuvres against opponents travelling up to 50%–70% of the ego vehicle’s speed, depending on conditions. The ego vehicle has a maximum speed of 2 m/s. Overall, the proposed system demonstrates that modular, classical architectures can achieve competitive, real-time performance in dynamic multi-vehicle racing. The results provide a foundation for future research in competitive autonomous racing and higher-speed manoeuvres.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Flood, Christopher Michael

Advisors dc:contributor.advisor

• Engelbrecht, H. A.
• Schoeman, J. C.

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/135988

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/135988