Vision-based Lateral Control of Vehicles for Road following Using Reinforcement Learning and Neural Networks

Reinforcement learning (RL) is the problem faced by a controller that must learn and improve behavior through trial-and-error interactions with a dynamic environment. It is completely an on-line procedure and thus requires no model of the plant nor of the environment. Hence, it seems ideal for vehicle control where uncertainty abounds in terms of the ever-changing vehicle dynamics during its continuing interaction with the road environment. Furthermore, it could also be used in conjunction with supervised learning in that the RL controller may further enhance the performance of an already trained supervised controller without the need for any further training data. This proposed reinforcement learning architecture is found to have a sound on-line learning control performance especially at high speed road following of a high curvature road. Both computer simulation and actual experiments on a test vehicle have been performed.