NVIDIA最新！GraspGen：基于扩散模型的六自由度抓取生成框架

Core Viewpoint - GraspGen framework addresses the challenge of generalization in 6-DOF grasping by modeling the grasp generation process as an iterative diffusion process, enhancing grasp generation capabilities through the DiffusionTransformer architecture and an efficient discriminator for sampling evaluation [2][21]. Group 1: Core Methodology - GraspGen models the 6-DOF grasp generation as a diffusion process in SE(3) space, utilizing Denoising Diffusion Probabilistic Model (DDPM) for faster computation and simpler implementation compared to traditional energy-based models [4]. - The framework employs PointTransformerV3 (PTv3) to convert unstructured point clouds into structured formats, reducing translation error by 5.3mm and improving recall rate by 4% compared to PointNet++ [4]. - The noise prediction network generates grasps through a 10-step denoising process, significantly fewer than the hundreds of steps required for image diffusion [5]. Group 2: Discriminator Innovations - GraspGen's discriminator innovatively reuses the generator's object encoder, reducing memory usage by 21 times compared to traditional methods [7]. - The discriminator is trained on a dataset generated by the generator, allowing it to better identify failure modes such as collisions and distant grasps, achieving an AUC of 0.947 compared to 0.886 when trained solely on offline data [16][21]. Group 3: Experimental Results - In single-object scenarios, GraspGen's precision-recall curve AUC exceeds baseline by 48% on the ACRONYM dataset, demonstrating the importance of the discriminator [10]. - In cluttered scenes, GraspGen achieves the highest task success rate and grasp success rate, outperforming Contact-GraspNet by 16.9% and M2T2 by 7.8% [13]. - Real robot experiments on the UR10 robotic arm show an overall success rate of 81.3% across various scenarios, significantly higher than M2T2 (28%) and AnyGrasp (17.6%) [19]. Group 4: Limitations and Future Directions - GraspGen shows limitations in performance on cubical objects and relies heavily on the quality of depth sensing and instance segmentation, with training requiring approximately 3,000 GPU hours [21].