Evolving a general electronic stability program for car simulated in TORCS

Abstract

We present an approach of evolving (via Genetic Programming, GP) the electronic stability program (ESP) of a car, realistically simulated in The Open Racing Car Simulator (TORCS). ESP is intended to assist the yaw rotation of an unstable (e.g., either understeering or oversteering) car in low-grip, slippery road conditions by applying a carefully-timed asymmetrical braking forces to its wheels. In the proposed approach, the amount of ESP-induced brake force is represented as an evolvable (via GP) algebraic function (brake force function BFF) of the values of parameters, pertinent to the state of the car, and their derivatives. In order to obtain a general BFF, i.e., a function that result in a handling of the car, that is better than that of non ESP car, for a wide range of conditions, we evaluate the evolving BFF in several fitness cases representing different combinations of surface conditions and speeds of the car. The experimental results indicate that, compared to the car without ESP, the best evolved BFF of ESP offers a superior controllability — in terms of both (i) a smaller deviation from the ideal trajectory and (ii) faster average speed on a wide range of track conditions (“icy”, “snowy”, “rainy” and “dry”) and traveling speeds. Presented work could be viewed as an attempt to contribute a new functionality in TORCS that might enrich the experience of gamers by the enhanced controllability of their cars in slippery road conditions. Also, the results could be seen as a step towards the verification of the feasibility of applying GP for automated, evolutionary development of ESP.