A new generic object recognition (GOR) method for robot perception is proposed in this paper, based on multi-feature fusion of two-dimensional (2D) and 3D scale invariant feature transform descriptors drawn from 2D images and 3D point clouds. The trained support vector machine is utilized to construct multi-category classifiers that recognize the objects. According to our results, this new GOR approach achieves higher recognition rates than classical methods tested, even when one has large intra-class variations, or high inter-class similarities of the objects. Simulation results demonstrate the effectiveness and efficiency of the proposed GOR approach.