Intelligent Decision-Making and Safety Control System for Autonomous Vehicles in Extreme Traffic Scenarios

Abstract

The multiple challenges of low adhesion roads, nonlinear dynamics and dynamic uncertainty present challenges for the decision making and control systems of autonomous vehicles in extreme traffic. This paper proposes a hierarchical coupling architecture: the upper layer uses a soft actor-critic (CSAC) algorithm with safety constraints using a control barrier function (CBF) to obtain important features of traffic participants via an attention mechanism and map the action to a safety set using control barrier function (CBF) quadratic programming; the lower layer integrates nonlinear model predictive control (NMPC) with constraints of the Dugoff tire model, considering how the vehicle's action would affect the tire's stability, and adds constraints on slip ratio and yaw rate penalties. The CBF safety filtering in the CSAC decision layer reduces the safety constraint violation rate; the peak lateral position error of NMPC during emergency lane changes on slippery surfaces is only 0.033 m, and the slip ratio constraint limits the sideslip angle to within 3.15° with no yaw rate overshoot. The proposed architecture can effectively enhance the decision-making intelligence and control safety of autonomous vehicles in extreme traffic environment.