Abstract Learning from human skills has become one of the popular inspirations in grasp prediction and evaluation, but lack of effective methods on groups of grasp points for multi-fingered dexterous hands yields an open challenge. When facing an object, humans firstly predict a variety of options for grasps, which can be concerned as a complex multi-valued problem. After prediction, humans evaluate grasps and then choose the optimal one. Inspired by human skills, we propose Grasp Prediction Networks (GPNs) based on Convolutional Neural Networks (CNNs) and Mixture Density Networks (MDNs). The proposed GPNs map from a depth image to a set of parameters for Gaussian Mixture Model (GMM), from which candidate groups of grasp points can be sampled for prediction. Besides, we also propose Grasp Evaluation Networks (GENs) to evaluate candidate groups and then choose the optimal group of grasp points. The proposed GENs consider force-closure metric as grasp quality for evaluation. Different from other related work, our method (1) utilizes a probabilistic model to predict multiple groups of grasp points from a monocular depth image and (2) evaluates grasp quality with force-closure metric given a monocular depth image and a group of grasp points. Furthermore, we built a grasp dataset which consists of depth images, groups of grasp points and each group’s grasp quality. Herein, three different experiments were designed to validate our approach. The first one was a comparative experiment and revealed that GPNs show equivalent performance as GraspIt! in terms of high-quality grasp planning. The second one was also a comparative experiment, which validated that GENs can evaluate grasps as precisely as GraspIt!. Moreover, the last one was an actual experiment implemented on Shadow Hand Lite, and experimental results indicated that our approach achieved finely grasp of novel objects.

中文翻译：

---

基于深度学习的多指灵巧手抓握预测与评估

摘要 从人类技能中学习已成为抓握预测和评估的流行灵感之一，但缺乏有效的多指灵巧手抓点组方法，这是一个开放的挑战。当面对一个物体时，人类首先预测各种抓取选项，这可以被视为一个复杂的多值问题。预测后，人类评估抓握，然后选择最佳抓握。受人类技能的启发，我们提出了基于卷积神经网络 (CNN) 和混合密度网络 (MDN) 的抓握预测网络 (GPN)。所提出的 GPN 从深度图像映射到高斯混合模型 (GMM) 的一组参数，可以从中抽取候选抓取点组进行预测。除了，我们还提出了抓取评估网络（GENs）来评估候选组，然后选择最佳抓取点组。提议的 GEN 将力闭合指标视为评估的抓取质量。与其他相关工作不同，我们的方法（1）利用概率模型从单眼深度图像中预测多组抓握点，（2）在给定单眼深度图像和一组抓握点的情况下，使用力闭合度量评估抓握质量. 此外，我们构建了一个抓取数据集，其中包含深度图像、抓取点组和每组抓取质量。在此，设计了三个不同的实验来验证我们的方法。第一个是比较实验，结果表明 GPN 表现出与 GraspIt 相同的性能！在高质量抓规划方面。第二个也是比较实验，它验证了 GEN 可以像 GraspIt! 一样精确地评估抓取。此外，最后一个是在 Shadow Hand Lite 上实现的实际实验，实验结果表明我们的方法实现了对新物体的精细把握。