Graspness Discovery in Clutters for Fast and Accurate Grasp Detection

TL;DR

"graspness", a quality based on geometry cues that distinguishes graspable area in cluttered scenes, is proposed and a look-ahead searching method is proposed for measuring the graspness and statistical results justify the rationality of the method.

Abstract

Efficient and robust grasp pose detection is vital for robotic manipulation. For general 6 DoF grasping, conventional methods treat all points in a scene equally and usually adopt uniform sampling to select grasp candidates. However, we discover that ignoring where to grasp greatly harms the speed and accuracy of current grasp pose detection methods. In this paper, we propose "graspness", a quality based on geometry cues that distinguishes graspable areas in cluttered scenes. A look-ahead searching method is proposed for measuring the graspness and statistical results justify the rationality of our method. To quickly detect graspness in practice, we develop a neural network named cascaded graspness model to approximate the searching process. Extensive experiments verify the stability, generality and effectiveness of our graspness model, allowing it to be used as a plug-and-play module for different methods. A large improvement in accuracy is witnessed for various previous methods after equipping our graspness model. Moreover, we develop GSNet, an end-to-end network that incorporates our graspness model for early filtering of low-quality predictions. Experiments on a large-scale benchmark, GraspNet-1Billion, show that our method outperforms previous arts by a large margin (30+ AP) and achieves a high inference speed. The library of GSNet has been integrated into AnyGrasp, which is at https://github.com/graspnet/anygrasp_sdk.

Figures (8)

• Figure 1: Graspness illustration for a cluttered scene. Brighter color denotes higher graspness. We prefer the points with high graspness for grasping.
• Figure 2: Graspness scores. The left image shows the graspness without collision detection while the right image shows the graspness with collision detection
• Figure 3: t-SNE visualization of encoded local geometry. Orange points denote the samples with high graspness, and blue points denote the samples with low graspness.
• Figure 4: GSNet architecture. The two rows show the process of cascaded graspness model and grasp operation model respectively. In cascaded graspness model, point encoder-decoder outputs $C$-dim feature vectors for the input $N$ points. A point-wise graspable landscape is generated and $M$ seed points are sampled from it. The seeds are then used to generate view-wise graspable landscapes, and select the grasp view. In grasp operation model, the seeds are grouped in cylinder regions. The grasp scores and gripper widths are predicted for each group and used to output $M$ grasp poses.
• Figure 5: Qualitative results of GSNet. Top 50 grasps after grasp-NMSgraspnet are displayed.