In this work, we develop a modular robotic system to predict single or multiple grasping poses specifically for parallel-plate robotic grippers using RGB and depth images. An end-to-end Inverted Residual Convolutional Neural Network (IR-ConvNet) model is proposed. The model uses state-of-the art Fused-MBconv blocks for feature extraction. Transposed Convolution is applied to up-sample the feature maps to thesize of the input. The grasp proposals can be inferred from the three-channel output feature maps. Given the efficient number of parameters, the model can run in real-time. The proposed model is evaluated on two public grasping datasets and a set of casual objects. The best model variant can achieve accuracy of 97.8%and 96.6% on image-wise splitting and object-wise splitting tests on Cornell Grasp Dataset respectively. The Jacquard Dataset accuracy is 93.9%. Finally, a robotic arm UR-5 is used to implement the detected grasps. The experimental results show effectiveness of the proposed robot grasping detection and implementation system.