Generating Complex Driving Behavior by Means of Neural Fields

We have developed a neural field based architecture to generate complete behavioral sequences for autonomous driving. On the basis of a user defined desired speed, the system can autonomously decide between the different behavioral alternatives lane following, lane change, acceleration and deceleration. Depending on the current traffic situation, the system can autonomously organize reactive behavioral sequences such as overtaking. Basically, the system consists of coupled one-dimensional neural fields: two fields for a desired longitudinal position and one field for the lateral position. All field dynamics are driven in a parameter regime which guarantees the existence of a monomodal self stabilizing peak. Both, the longitudinal and the lateral control consist of two dynamics: one dynamics represents a planning component (“motivation”) and one determining the action component. The action level generates the desired position of movement which determines the corresponding controlled variables: steering angle and velocity.