Real-Time Collision Imminent Steering Using One-Level Nonlinear Model Predictive Control

Abstract

Automotive active safety features are designed to complement or intervene a human driver's actions in safety critical situations. Existing active safety features, such as adaptive cruise control and lane keep assist, are able to exploit the ever growing sensor and computing capabilities of modern automobiles. An emerging feature, collision imminent steering, is designed to perform an evasive lane change to avoid collision if the vehicle believes collision cannot be avoided by braking alone. This is a challenging maneuver, as the expected highway setting is characterized by high speeds, narrow lane restrictions, and hard safety constraints. To perform such a maneuver, the vehicle may be required to operate at the nonlinear dynamics limits, necessitating advanced control strategies to enforce safety and drivability constraints.

This dissertation presents a one-level nonlinear model predictive controller formulation to perform a collision imminent steering maneuver in a highway setting at high speeds, with direct consideration of safety criteria in the highway environment and the nonlinearities characteristic of such a potentially aggressive maneuver. The controller is cognizant of highway sizing constraints, vehicle handling capability and stability limits, and time latency when calculating the control action. In simulated testing, it is shown the controller can avoid collision by conducting a lane change in roughly half the distance required to avoid collision by braking alone. In preliminary vehicle testing, it is shown the control formulation is compatible with the existing perception pipeline, and prescribed control action can safely perform a lane change at low speed.

Further, the controller must be suitable for real-time implementation and compatible with expected automotive control architecture. Collision imminent steering, and more broadly collision avoidance, control is a computationally challenging problem. At highway speeds, the required time for action is on the order of hundreds of milliseconds, requiring a control formulation capable of operating at tens of Hertz. To this extent, this dissertation investigates the computational expense of such a controller, and presents a framework for designing real-time compatible nonlinear model predictive controllers. Specifically, methods for numerically simulating the predicted vehicle response and response sensitivities are compared, their cross interaction with trajectory optimization strategy are considered, and the resulting mapping to a parallel computing hardware architecture is investigated. The framework systematically evaluates the underlying numerical optimization problem for bottlenecks, from which it provides alternative solutions strategies to achieve real-time performance. As applied to the baseline collision imminent steering controller, the procedure results in an approximate three order of magnitude reduction in compute wall time, supporting real-time performance and enabling preliminary testing on automotive grade hardware.