This paper proposes a method for picking a specified number of pieces of loose-packed French fries using a robotic arm. The proposed method consists of a two-stage process: first, estimating the number of pieces to be grasped for each grasping action using a deep learning model; and second, determining the optimal grasping action based on this estimation. The first-stage grasping quantity estimation model accounts for variation in grasping results and outputs the number of pieces as a probability distribution. In the second stage, it selects and performs grasping at a point that is close to the specified number and has low prediction uncertainty. To verify the effectiveness of the proposed method, experiments were conducted using food samples of French fries placed in a ball container. These experiments evaluated the feasibility of grasping the specified number under three different bulk states.