[論文レビュー] Dexterous grasp data augmentation based on grasp synthesis with fingertip workspace cloud and contact-aware sampling

Liqi Wu, Haoyu Jia+3 more|arXiv (Cornell University)|Mar 17, 2026

Robot Manipulation and Learning被引用数 0

ひとこと要約

The paper introduces AutoWS and FSG to automatically generate fingertip workspace clouds and fast, contact-aware grasp poses for dexterous hands, enabling efficient data augmentation for data-driven grasp training.

ABSTRACT

Robotic grasping is a fundamental yet crucial component of robotic applications, as effective grasping often serves as the starting point for various tasks. With the rapid advancement of neural networks, data-driven approaches for robotic grasping have become mainstream. However, efficiently generating grasp datasets for training remains a bottleneck. This is compounded by the diverse structures of robotic hands, making the design of generalizable grasp generation methods even more complex. In this work, we propose a teleoperation-based framework to collect a small set of grasp pose demonstrations, which are augmented using FSG--a Fingertip-contact-aware Sampling-based Grasp generator. Based on the demonstrated grasp poses, we propose AutoWS, which automatically generates structured workspace clouds of robotic fingertips, embedding the hand structure information directly into the clouds to eliminate the need for inverse kinematics calculations. Experiments on grasping the YCB objects show that our method significantly outperforms existing approaches in both speed and valid pose generation rate. Our framework enables real-time grasp generation for hands with arbitrary structures and produces human-like grasps when combined with demonstrations, providing an efficient and robust data augmentation tool for data-driven grasp training.

研究の動機と目的

• Motivate data-efficient generation of grasp poses for dexterous hands with varying structures.
• Embed hand anatomy into fingertip workspace representations to avoid inverse kinematics

提案手法

• AutoWS automatically generates fingertip workspace clouds by embedding finger structure into workspace representations.
• FSG (Fingertip-contact-aware Sampling-based Grasp generator) rapidly samples grasp poses using fingertip workspace clouds and object point clouds.
• A contact-face approximation reconstructs fingertip contact regions and enables joint-angle retrieval without IK.
• Grasp quality is evaluated with Grasp Wrench Space (GWS) epsilon to filter and rank candidates.
• Finger iteration allows varying the number of fingers used for a grasp and can disable less important fingers to speed generation.
• Link-penetration checking can be added to further validate feasible grasps.

実験結果

リサーチクエスチョン

• RQ1 Can AutoWS and FSG generate high-quality grasp poses for dexterous hands with arbitrary structures in real time?
• RQ2 Does fingertip-contact-aware sampling improve the speed and validity of generated grasps compared to baselines?
• RQ3 How does the proposed pipeline perform for data augmentation versus hand-only grasp data generation across diverse objects?

主な発見

Object	Method	Time Cost / Result (s)	GWS Epsilon	GWS Volume	Valid Epsilon Percentage
004-sugar-box	EigenGrasp	1.622	9.16e-03	8.96e-04	0.3%
004-sugar-box	DexGraspNet	0.525	1.62e-02	2.40e-03	1.5%
004-sugar-box	QD-Grasp-6DoF	4.225	9.58e-03	1.74e-03	6.4%
004-sugar-box	FSG	0.147	1.80e-02	2.13e-03	18.6%
004-sugar-box	FSG+Aff	0.077	2.10e-02	2.37e-03	19.1%
004-sugar-box	FSG+Aff-FinIter	0.187	2.18e-02	3.82e-03	16.4%
004-sugar-box	FSG+Aff+LinkPen	0.087	1.89e-02	2.18e-03	17.7%
005-tomato-soup-can	EigenGrasp	1.619	2.32e-02	8.52e-03	0.1%
005-tomato-soup-can	DexGraspNet	0.525	1.94e-02	1.09e-02	5.9%
005-tomato-soup-can	QD-Grasp-6DoF	13.847	2.07e-02	8.48e-03	36.5%
005-tomato-soup-can	FSG	0.207	2.90e-02	1.05e-02	34.9%
005-tomato-soup-can	FSG+Aff	0.056	2.64e-02	5.37e-03	60.8%
005-tomato-soup-can	FSG+Aff-FinIter	0.541	3.29e-02	1.01e-02	86.8%
005-tomato-soup-can	FSG+Aff+LinkPen	0.066	2.57e-02	5.32e-03	71.9%
024-bowl	EigenGrasp	1.621	NaN	NaN	0.0%
024-bowl	DexGraspNet	0.525	NaN	NaN	0.0%
024-bowl	QD-Grasp-6DoF	0.297	6.59e-03	3.11e-03	0.3%
024-bowl	FSG	NaN	NaN	NaN	NaN
024-bowl	FSG+Aff	NaN	NaN	NaN	NaN
024-bowl	FSG+Aff-FinIter	NaN	NaN	NaN	NaN
024-bowl	FSG+Aff+LinkPen	NaN	NaN	NaN	NaN
043-phillips-screwdriver	EigenGrasp	1.614	1.78e-02	5.39e-04	0.3%
043-phillips-screwdriver	DexGraspNet	0.525	2.65e-03	1.79e-04	0.2%
043-phillips-screwdriver	QD-Grasp-6DoF	0.579	5.40e-03	5.25e-04	2.2%
043-phillips-screwdriver	FSG	0.181	8.41e-03	4.38e-04	18.0%
043-phillips-screwdriver	FSG+Aff	0.128	6.38e-03	1.63e-04	36.2%
043-phillips-screwdriver	FSG+Aff-FinIter	2.536	1.22e-02	7.58e-04	45.0%
043-phillips-screwdriver	FSG+Aff+LinkPen	0.120	7.44e-03	2.19e-04	39.3%
062-dice	EigenGrasp	1.624	NaN	NaN	0.0%
062-dice	DexGraspNet	0.525	NaN	NaN	0.0%
062-dice	QD-Grasp-6DoF	2.719	NaN	NaN	0.0%
062-dice	FSG	NaN	NaN	NaN	NaN
062-dice	FSG+Aff	NaN	NaN	NaN	NaN
062-dice	FSG+Aff-FinIter	NaN	NaN	NaN	NaN
062-dice	FSG+Aff+LinkPen	NaN	NaN	NaN	NaN

• FSG significantly reduces grasp pose generation time (e.g., 0.147 s to generate poses for some cases) compared with baselines like EigenGrasp (1.622 s) and QD-Grasp-6DoF (4.225 s).
• FSG and its variants achieve higher valid epsilon percentages than baselines on several YCB objects (e.g., 18.6% for FSG vs 0.3% for EigenGrasp on 004-sugar-box).
• FSG+Aff further improves speed and validity by using the target grasp part cloud (e.g., 19.1% valid epsilon for 004-sugar-box).
• Finger iteration allows grasp generation with fewer fingers if needed, maintaining or improving feasibility.
• The method supports arbitrary hand structures and demonstrates generated grasps for multiple hands (Shadow Hand, Robotiq 2F-85, Barrett Hand, Allegro Hand).
• For cluttered or very small/thin objects, valid grasp poses may be limited by the workspace cloud and contact-face sampling.