Autonomous racing platforms such as the 1:10 scale RoboRacer serve as crucial testbeds for evaluating high-speed navigation algorithms, but have traditionally relied almost exclusively on 2D lidar for localization, obstacle detection, and control. Lidar provides highly accurate distance measurements and works reliably at night, but its flat planar scan limits the vehicle's ability to perceive complex 3D track geometry and degrades on reflective or transparent surfaces, bounding the performance of both classic control pipelines and data-driven reinforcement learning (RL) agents. Cameras can fill this gap with richer 3D and semantic information, but training and validating a fused sensor pipeline requires simulators capable of generating realistic camera data alongside lidar — a standard 2D simulator (e.g. the RoboRacer Gym) allows for rapid training but cannot simulate cameras, while a 3D simulator (e.g. AutoDrive) supports full stereo vision at higher computational cost.
The aim of this project is to develop and validate a lidar-camera fusion architecture for autonomous racing, and to evaluate how this sensor choice interacts with the design of the driving pipeline itself. The student will implement the fused perception stack under both a classical control-theoretic pipeline and an RL-based pipeline, contrasting model-based and learned approaches to using the fused sensor data. The task is single-vehicle time-trial racing, allowing the project to isolate questions of sensor fusion and pipeline design from the added complexity of multi-agent interaction. The fused pipeline will be evaluated on the physical track against a lidar-only baseline, using metrics including localization/perception accuracy, robustness to lighting and reflective surfaces, and lap-time performance.