An Exploration of Domain Adaptation Applying to Grasp Detection Algorithm

Grasping is one of the main approaches for robots to move and manipulation objects. Recently, more and more deep neural networks which need a large amount of data are applied to machine learning. However, the number of images in the real scene is far less than the number of images in the datasets commonly used in deep learning. And there is a certain difference between the simulation data and real data. So there will be a phenomenon of domain migration, which will lead to a decline in network performance. In order to make the network more efficient and extract more accurate features, we design and implement the algorithm combining MMD oriented anchor frame mechanism to capture and detect the grasp location. Different from prior research has directly reducing the MMD distance between source domain and target domain , we between source domain and target domain to produce a Gaussian distribution as the middle distribution to reduce the gap between simulation data and real data. After experimental verification, the accuracy of our algorithm on the simulation dataset and real dataset reached 90.10% and 90.96% respectively.