MTGrasp: Multiscale 6-DoF Robotic Grasp Detection

The actual grasping scenarios are often complex, and achieving reliable and stable robot grasping in complex scenes is an important problem. Due to its high flexibility, 6-DoF grasping methods are well-suited for representing grasping poses in complex scenes. In order to achieve more accurate 6-DoF grasping, this article proposes a new multiscale 6-DoF grasp pose detection method. First, based on the object model, we can generate multiple grasp poses on the object model. Evaluating the quality of the grasps in a reasonable and effective way is a key issue. To achieve more accurate grasp pose evaluation, we propose a multimetric grasp pose evaluation algorithm to assess grasp poses more accurately and assign a grasp quality score. The neural network learns and detects the best grasp poses based on the grasp quality score. Next, considering the existence of objects at multiple scales in practical scenes, we propose a multiscale grasp pose search algorithm to detect grasp candidate regions and search for multiscale grasp poses. Based on the proposed algorithms, we construct a new 6-DoF grasp pose detection network called MTGrasp and train/test it on a large-scale grasp dataset, GraspNet-1Billion dataset. Experimental results show that our proposed algorithm achieves state-of-the-art performance on the dataset. Finally, we apply MTGrasp to actual grasping experiments with the UR3 robot, resulting in a higher grasp success rate.