Optimized sensory-motor couplings plus strategy extensions for the TORCS car racing challenge

Abstract

The TORCS simulated car racing competition was introduced over a year ago. It asks for the design of racing car control strategies that have to rely on local track and driving information only, such as distance sensors, angle-to-track axis, or velocity vectors. Thus, the competition asks for strategies that are sensory-motorically grounded rather than strategies that can be designed (online or even offline) by an external observer that has full track knowledge. Moreover, the competition enforces the development of rather general driving strategies since optimization is on driving success in general rather than driving success on one particular track. This paper describes the steps taken to develop COBOSTAR, an autonomous racing car strategy with several general, context-dependent behavioral modules and strategic advancements. Most of the behavioral parameters were optimized with covariance matrix adaptation evolutionary strategy techniques. COBOSTAR won the simulated car racing competition at the IEEE Congress of Evolutionary Computation (CEC 2009) and there is still lots of room for further optimizations and strategy additions. Apart from describing the COBOSTAR racer in detail, we also outline possible next steps and future challenges.