How to detect the object and guide the robot to get close to the object is an important task for autonomous robots. The main difficulties here is that the view of the robot changes a lot when it moves and there are limited data available to train. To tackle these challenges, we propose a novel vision system for the robot, the model adaption object detection system. Instead of using one object detection neural network to solve all the problem, we use different object detection neural network to guide the robot according to the situation the robot is in, by using a meta neural network to allocate the object detection neural network. Furthermore, we use the transfer learning technology and depthwise separable convolutions, so that our model is easy to train and can address small dataset problem.