F1tenth

Mapping and gap-following produce a seed path during track exploration; a raceline generator refines this path into an initial optimal trajectory. The primary controller uses Frenet-frame MPC with a kinematic bicycle model, curvature-based speed estimation, and a confidence metric adapted from lap-to-lap tracking error. This design supports fast, reliable operation in known, high-friction environments. A sampler-based MPCC, inspired by MPPI, remains under development, using a dynamic bicycle model, RK4 forward rollouts, a 0.8-second prediction horizon, and contouring and progress costs that account for obstacle interactions, supporting more capable decisions under slip and around obstacles. Feel free to reach out to me for additional details.

Obstacle detection begins by filtering LiDAR measurements with a two-state Kalman filter to detect nearby opponents and estimate their velocity. Gaussian regression then refines the predicted opponent trajectory, allowing an optimal overtake point to be selected; a rerouting stage modifies the reference path for the MPC. The sampler-based controller handles this more directly by transforming LiDAR observations into the map frame and incorporating an additional LiDAR-derived cost, making obstacle awareness inherent to trajectory evaluation. Contouring costs further balance progress, rejoining the optimal path, and contouring error during obstacle avoidance.