Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge

TL;DR

The paper tackles the generalization challenge of learning-based 6-DoF grasp detection to unseen objects by introducing domain priors. It presents a two-pronged framework: Physical Constraint Regularization (PCR), which enforces physical rules through a differentiable Signed Distance Function, and Contact-Score Joint Optimization (C-SJO), which refines unstable grasps at test time using a projection contact map and a grasp-score prior. Key contributions include integrating antipodal, surface, and collision constraints into end-to-end learning with differentiable geometry, and a test-time refinement strategy that improves performance in cluttered scenes. Experiments on GraspNet-1billion and real-robot trials demonstrate substantial gains on novel objects and robust real-world grasping, reducing reliance on extensive data augmentation.

Abstract

We focus on the generalization ability of the 6-DoF grasp detection method in this paper. While learning-based grasp detection methods can predict grasp poses for unseen objects using the grasp distribution learned from the training set, they often exhibit a significant performance drop when encountering objects with diverse shapes and structures. To enhance the grasp detection methods' generalization ability, we incorporate domain prior knowledge of robotic grasping, enabling better adaptation to objects with significant shape and structure differences. More specifically, we employ the physical constraint regularization during the training phase to guide the model towards predicting grasps that comply with the physical rule on grasping. For the unstable grasp poses predicted on novel objects, we design a contact-score joint optimization using the projection contact map to refine these poses in cluttered scenarios. Extensive experiments conducted on the GraspNet-1billion benchmark demonstrate a substantial performance gain on the novel object set and the real-world grasping experiments also demonstrate the effectiveness of our generalizing 6-DoF grasp detection method.

Figures (7)

• Figure 1: Illustration of (a) the performance and distribution gap between objects similar to training samples (similar objects) and those with largely varied shapes and structures (novel objects) and (b) the limitations of the object augmentation methods.
• Figure 2: The pipeline of differential physical constraint integration. With the fusion point-cloud, the differentiable 6-DoF grasp network predicts the grasp configurations. The position of contacts $c_1,c_2$ are calculated from grasp configurations by the gripper model and the regularization $R$ is computed from the object SDF for back-propagation.
• Figure 3: (a) Calculation of the contact map and (b) comparison of the projection contact map and the original version.
• Figure 4: The pipeline of contact optimization and score optimization with ContactNet and ScoreNet.
• Figure 5: Visualization of the process of C-SJO.