CAPGrasp: An $\mathbb{R}^3\times \text{SO(2)-equivariant}$ Continuous Approach-Constrained Generative Grasp Sampler

TL;DR

Overall, CAPGrasp is a sample-efficient solution when grasps must originate from specific directions, such as grasping in confined spaces, and more than three times as sample efficient as unconstrained grasp samplers.

Abstract

We propose CAPGrasp, an $\mathbb{R}^3\times \text{SO(2)-equivariant}$ 6-DoF continuous approach-constrained generative grasp sampler. It includes a novel learning strategy for training CAPGrasp that eliminates the need to curate massive conditionally labeled datasets and a constrained grasp refinement technique that improves grasp poses while respecting the grasp approach directional constraints. The experimental results demonstrate that CAPGrasp is more than three times as sample efficient as unconstrained grasp samplers while achieving up to 38% grasp success rate improvement. CAPGrasp also achieves 4-10% higher grasp success rates than constrained but noncontinuous grasp samplers. Overall, CAPGrasp is a sample-efficient solution when grasps must originate from specific directions, such as grasping in confined spaces.

Figures (7)

• Figure 1: An example of picking from a shelf: generating grasps with specified approach constraints and choosing the highest scoring one (red) for execution.
• Figure 2: A visualization of the problem statement. $\mathbf{O}$ is represented by the blue point cloud, while $\mathbf{A}$, parameterized by $\Vec{v}_{A}$ and $\alpha$, is visualized as the purple cone. The goal is to generate a grasp pose $\Vec{g}$ whose unit approach vector $\Vec{v}_{g}$ lies inside $\mathbf{A}$.
• Figure 3: CAPGrasp grasp sampling. (a) $\mathbf{O}$ and $\mathbf{A}$ are in the camera space. (b) $\mathbf{O}$ and $\mathbf{A}$ are transformed into the approach space using the transformation $\mathbf{T_{v_{A}}}$ that aligns $\Vec{v}_A$ with $-\Vec{y}$. (c) CAPGrasp generates grasps $\mathbf{\hat{G}}$ whose approach directions are within $\alpha$. (d) $\mathbf{\hat{G}}$ is transformed back to the camera space using $\mathbf{T_{v_{A}}^{-1}}$.
• Figure 4: The training method for CAPGrasp. (a) Sample an object-grasp pair in the camera space. (b) A random $\alpha_{i}$ is sampled (Stage 1 in \ref{['alg:training_alg']}). The resulting blue cone, parameterized by $\Vec{v}_{g}$ and $\alpha_{i}$, represents the space from which we can sample $\Vec{v}_A$. (c) $\Vec{v}_{A}$ is sampled inside the blue cone (Stage 2 in \ref{['alg:training_alg']}). The pink cone, parameterized by $\Vec{v}_{A}$ and $\alpha_{i}$, now represents $\mathbf{A}$. (d) The object-grasp pair is transformed into the approach space and becomes the conditional training pair.
• Figure 5: The success-over-coverage for GoNet with different yaw/pitch resolutions, CAPGrasp with different $\alpha$, and GraspNet. The auc score for each method is visualized in the image legend.