The paper presents a deep learning approach to detect robotic grasps in RGB-D images, addressing the challenges of evaluating a large number of candidate grasps and handling multimodal inputs effectively. The authors propose a two-step cascaded system with two deep networks: the first network uses fewer features to quickly prune out unlikely grasps, while the second network with more features refines the top detections. They also introduce a structured regularization method based on multimodal group regularization to improve the quality of features learned from RGB-D data. The method is evaluated on a challenging dataset and shown to outperform state-of-the-art methods for rectangle-based grasp detection, achieving high success rates on both Baxter and PR2 robots. The contributions include a deep learning algorithm for robotic grasp detection, a new method for handling multimodal inputs, and a multi-step cascaded detection system that reduces computational cost while improving performance.The paper presents a deep learning approach to detect robotic grasps in RGB-D images, addressing the challenges of evaluating a large number of candidate grasps and handling multimodal inputs effectively. The authors propose a two-step cascaded system with two deep networks: the first network uses fewer features to quickly prune out unlikely grasps, while the second network with more features refines the top detections. They also introduce a structured regularization method based on multimodal group regularization to improve the quality of features learned from RGB-D data. The method is evaluated on a challenging dataset and shown to outperform state-of-the-art methods for rectangle-based grasp detection, achieving high success rates on both Baxter and PR2 robots. The contributions include a deep learning algorithm for robotic grasp detection, a new method for handling multimodal inputs, and a multi-step cascaded detection system that reduces computational cost while improving performance.