Abstract-Much research has been done on the subject of collision avoidance (COLAV). However, few results are presented that consider vehicles with second-order nonholonomic constraints, such as autonomous underwater vehicles (AUVs). This paper considers the dynamic window (DW) algorithm for reactive horizontal COLAV for AUVs, and uses the HUGIN 1000 AUV in a case study. The DW algorithm is originally developed for vehicles with first-order nonholonomic constraints and is hence not directly applicable for AUVs without resulting in degraded performance. This paper suggests further developments of the DW algorithm to make it better suited for use with AUVs. In particular, a new method for predicting AUV trajectories using a linear approximation which accounts for second-order nonholonomic constraints is developed. The new prediction method, together with a modified search space, reduces the mean square prediction error to about one percent of the original algorithm. The performance and robustness of the modified DW algorithm is evaluated through simulations using a nonlinear model of the HUGIN 1000 AUV.