Autonomous racing provides a demanding benchmark for perception, planning and control algorithms : a race car must operate at the edge of its physical performance envelope, where nonlinear tire dynamics impose extreme requirements on each sub-system. Solving this problem requires constructing a map of the environment, estimating the vehicle state, generating an optimal trajectory, and tracking it under strict dynamic constraints, but the systematic integration of these sub-systems on a unified platform remains an open challenge.
This thesis focuses on trajectory planning and tracking for the F1TENTH autonomous racing platform. Experimentation is done entirely in simulation through F1TENTH Gym.
An offline trajectory planner is introduced which provides both a racing line with a minimum curvature and a velocity profile. The geometric trajectory is computed through solving an optimization problem under the form of a quadratic program. The velocity profile is obtained with a forward backward integration scheme that respects the friction circle constraints.
Two types of controllers are explored. The first one is a model-based implemented as a Model Predictive Controller (MPC) where a constrained non-linear program is solved over a receding finite horizon. The second one is a data-driven. The Soft Actor-Critic (SAC) algorithm is used to train reinforcement learning agents as control policies. These policies are trained without depending on a reference trajectory.
Dang, V. (Author),
Belleau, C. (Supervisor) &
Roberge, J.-P. (Co-supervisor),
8 Oct 2026Student thesis: Master's thesis › Master in Engineering: Information Technology Engineering