UGG: Unified Generative Grasping (2311.16917v2)

Abstract: Dexterous grasping aims to produce diverse grasping postures with a high grasping success rate. Regression-based methods that directly predict grasping parameters given the object may achieve a high success rate but often lack diversity. Generation-based methods that generate grasping postures conditioned on the object can often produce diverse grasping, but they are insufficient for high grasping success due to lack of discriminative information. To mitigate, we introduce a unified diffusion-based dexterous grasp generation model, dubbed the name UGG, which operates within the object point cloud and hand parameter spaces. Our all-transformer architecture unifies the information from the object, the hand, and the contacts, introducing a novel representation of contact points for improved contact modeling. The flexibility and quality of our model enable the integration of a lightweight discriminator, benefiting from simulated discriminative data, which pushes for a high success rate while preserving high diversity. Beyond grasp generation, our model can also generate objects based on hand information, offering valuable insights into object design and studying how the generative model perceives objects. Our model achieves state-of-the-art dexterous grasping on the large-scale DexGraspNet dataset while facilitating human-centric object design, marking a significant advancement in dexterous grasping research. Our project page is https://jiaxin-lu.github.io/ugg/.

Overview of the Unified Generative Grasping Model

The development of dexterous grasping in robotics focuses on creating a variety of successful grasping postures. Two principal methods have dominated this area: regression-based methods, which predict specific grasp parameters and often lack diversity, and generation-based methods, which excel at variety but sometimes fall short on success rates. The Unified Generative Grasping (UGG) model undertakes the challenge of combining these two approaches to leverage their advantages and mitigate their shortcomings.

Designing the UGG Model

The UGG model is built upon a foundation that bridges object, hand, and contact point information seamlessly within a unified diffusion model. This model applies an all-transformer architecture and introduces a novel representation called Contact Anchors. Contact Anchors provide a concise representation for contact points which is critical for successful grasping and simultaneously compatible with the unified generative process. The model operates in the latent space of the object's point cloud and the hand's parameters.

Improving Grasp Quality and Diversity

An essential aspect of the UGG model is the inclusion of a lightweight discriminator that uses simulated data to enhance the model's grasping success rate while maintaining high diversity. This discriminator selects among generated grasps, favoring those more likely to succeed in achieving dexterous grasping.

Expanding Capabilities Beyond Grasp Generation

The UGG model's capabilities extend into human-centric object design. Not only can it generate grasps for fixed objects, but it can also create objects and contact points from given hand poses. This bi-directional generative ability offers valuable insights into designing objects that consider the nuances of hand-object interactions. These insights drive further research in object representation within robotic applications and contribute to the development of human-centric design guidelines.

Performance and Contributions

UGG demonstrates its effectiveness on the significant DexGraspNet dataset, setting a new benchmark for dexterous grasping research. Its unified approach not only streamlines the process of dexterous grasp generation but also opens up new possibilities in object recognition and robotic applications by offering an adaptable framework for a variety of tasks in the manufacturing, healthcare, agriculture, and extended reality domains.

In summary, UGG represents a significant leap in dexterous robotic grasping, marrying high success rates and diversity, while also presenting a flexible generative model for hand-object interaction tasks.