Learning-based grasp pose detection algorithms have boosted the performance of robot grasping, but they usually need manually fine-tuning steps to find the balance between detection accuracy and efficient. In this paper, we discard these intermediate procedures, like sampling grasps and generating grasp proposals, and propose an end-to-end grasp pose detection model. Our model uses the RGB image as the input and predicts the single grasp pose in each small grid of the image. Furthermore, the best grasps are found by non-maximum suppression (NMS) strategy. The clustering and ranking procedures are left for NMS while the network only generates dense grasp predictions, which keeps the network simple and efficient. To achieve dense predictions, the predicted grasps of our detection model are represented by the 6 channels images with each pixel location representing a rated grasp. To the best of our knowledge, our model is the first neural network that attaches a grasp pose in pixel level. The model achieves 96.5% accuracy which costs 14ms for prediction of a 480×360 resolution RGB image in Cornell Grasp Dataset, and 90.4% robot grasping success rate for unknown objects with a parallel plate gripper in the real environment.
Paper
Full text
High Accuracy and Efficiency Grasp Pose Detection Scheme with Dense Predictions
OpenAlex · Robot Manipulation and Learning · 2020
Abstract
Learning-based grasp pose detection algorithms have boosted the performance of robot grasping, but they usually need manually fine-tuning steps to find the balance between detection accuracy and efficient. In this paper, we discard these intermediate procedures, like sampling grasps and generating grasp proposals, and propose an end-to-end grasp pose detection model. Our model uses the RGB image as the input and predicts the single grasp pose in each small grid of the image. Furthermore, the best grasps are found by non-maximum suppression (NMS) strategy. The clustering and ranking procedures are left for NMS while the network only generates dense grasp predictions, which keeps the network simple and efficient. To achieve dense predictions, the predicted grasps of our detection model are represented by the 6 channels images with each pixel location representing a rated grasp. To the best of our knowledge, our model is the first neural network that attaches a grasp pose in pixel level. The model achieves 96.5% accuracy which costs 14ms for prediction of a 480×360 resolution RGB image in Cornell Grasp Dataset, and 90.4% robot grasping success rate for unknown objects with a parallel plate gripper in the real environment.