In this paper, an imitation learning approach of vision guided reaching skill is proposed for robotic precision manipulation, which enables the robot to adapt its end-effector’s nonlinear motion with the awareness of collision-avoidance. The reaching skill model firstly uses the raw images of objects as inputs, and generates the incremental motion command to guide the lower-level vision-based controller. The needle’s tip is detected in image space and the obstacle region is extracted by image segmentation. A neighborhood-sampling method is designed for needle component collision perception, which includes a neural networks based attention module. The neural network based policy module infers the desired motion in the image space according to the neighborhood-sampling result, goal and current positions of the needle’s tip. A refinement module is developed to further improve the performance of the policy module. In three dimensional (3D) manipulation tasks, typically two cameras are used for image-based vision control. Therefore, considering the epipolar constraint, the relative movements in two cameras’ views are refined by optimization. Experimental are conducted to validate the effectiveness of the proposed methods.