This talk describes an ongoing project aimed at developing a rigorous computational model of driver behavior in the ACT-R cognitive architecture. Driving is a complex dynamic task that requires the management and integration of many cognitive and visual-motor skills. For the purposes of operational control (e.g., steering), these skills include the encoding of relevant perceptual variables and the updating of continuous control signals through motor actions. For the purposes of higher-level processing, these skills include maintenance of situation awareness and decision making that acts upon this awareness.

We are developing an ACT-R model that accounts for behavior in the specific domain of highway driving (Salvucci, Boer, & Liu, submitted). The model is currently able to navigate a four-lane highway and maintain lane position on straight segments and curves. Control is performed through iterative encoding of the “near” and “far” regions of the roadway (Donges, 1978; Land & Horwood, 1995; Land & Lee, 1994) and discrete updates to steering based on aspects of these two regions (e.g., the vanishing point or tangent point as the “far” point). The model also maintains awareness of the surrounding vehicles and changes lanes when appropriate. Such awareness arises from occasional monitoring of other vehicles and updating of the “mental model” that tracks them in declarative memory and can predict their current location based on previous encodings. Decisions to change lanes are made on a satisficing basis when the car is too close to the lead car and there is adequate space in the other lane.

We validate the model by comparing its behavior to that of human drivers in our fixed-based driving simulator. In an empirical study (Salvucci & Liu, in preparation), ten drivers navigated a four-lane highway (two lanes in each direction) with moderate traffic generated by autonomous vehicles. The data collected in the study included standard control data (e.g., steering, acceleration, braking) as well as driver eye movements. We compared driver behavior to model behavior by focusing on steering and eye movements during the tasks of curve negotiation and lane changing. We find that the model nicely captures a number of the phenomena exhibited by human drivers (in our study and others), including the ratio of gazes to different regions of the road and to other vehicles, shifts of attention to the destination lane just at the start of a lane change, and preparatory steering before entering a curve.

The talk will provide an overview of this model to provide a basic understanding of its operation. In addition, the talk will outline some preliminary ideas about how the model facilitates real-time inference of driver intentions in the implementation of intelligent vehicle systems (see also Salvucci, 2000).

Acknowledgments

This work is being done in collaboration with Erwin Boer of Nissan Cambridge Basic Research and Andrew Liu of the MIT Man Vehicle Laboratory.

References

Anderson, J. R., and Lebiere, C. (1998). The Atomic Components of Thought. Hillsdale, NJ: Erlbaum.

Donges, E. (1978). A two-level model of driver steering behavior. Human Factors, 20, 691-707.

Land, M., & Horwood, J. (1995). Which parts of the road guide steering?. Nature, 377, 339-340.

Land, M. F., & Lee, D. N. (1994). Where we look when we steer. Nature, 369, 742-744.

Salvucci, D. D. (2000). Cognitive architectures for intelligent support systems. To appear in Proceedings of the ACT-R Workshop, August 5-7, 2000.

Salvucci, D. D., Boer, E. R., & Liu, A. (submitted). Modeling driver behavior in a cognitive architecture.

Salvucci, D. D., & Liu, A. (in preparation). Control and monitoring in highway driving.