Title: The Edge of Disaster: A Battle Between Autonomous Racing and Safety

Abstract: Autonomous racing represents a uniquely challenging control environment where agents must act while on the limits of a vehicle's capability in order to set competitive lap times. This places the agent on a knife's edge, with a very small margin between success and loss of control. Pushing towards this limit leads to a practical tension: we want agents to explore the limitations of vehicle control to maximise speed, but inadvertently going past that limit and losing control can cause irreparable damage to the vehicle itself. We provide a model predictive control (MPC) baseline that is able to, in a single lap, safely adapt to an unseen racetrack and achieve competitive lap times. Our approaches efficacy is demonstrated in simulation using the Learn To Race Challenge's environment and metrics.