In context
In early 2023, service robots were increasingly expected to perform dexterous manipulation in unstructured human environments. While two-finger grippers dominated industrial automation, multi-fingered dexterous hands offered greater flexibility for complex objects but faced high-dimensional grasp planning challenges. This paper from the University of Science and Technology of China addressed a key gap: efficiently optimizing grasping gestures for unknown objects without precise 3D models.
What was reported
Researchers developed a grasping gesture optimization method based on a convolutional neural network (CNN). They first constructed a grasp dataset in simulation (GraspIt and Gazebo) containing 3,089 samples, each with object depth images, hand configurations, and force-closure quality metrics. A grasping quality evaluation network (GQEN) was trained to predict the force-closure metric from monocular visual information and hand pose, achieving evaluation errors within 6%.
The grasp planning problem was reformulated as maximizing the predicted grasp quality. Using backpropagation and gradient ascent with backtracking line search, the method iteratively refined an initial suboptimal gesture to reach a local optimum, respecting joint limits. This approach avoided the multi-valued mapping issue common in regression-based methods.
Experiments in simulation compared GQEN evaluations with GraspIt, showing reliable quality prediction. On a real Shadow Hand Lite platform, the method achieved a grasping success rate above 80% for unknown objects. For initially failed grasps, optimization improved success in 90% of cases.
Why it mattered
This work demonstrated a practical, learning-based path to dexterous manipulation for unknown objects, potentially extending robotic automation beyond rigid two-finger tasks. By enabling efficient gesture optimization without full object models, it pointed toward more adaptive and versatile robotic hands in service and industrial settings.
“The results show that the grasping success rate of the proposed method for the unknown objects is more than 80%, and for the failed grasps, the success rate after optimization reaches 90%.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2023-03-06 · “基于深度神经网络的多指灵巧手抓取手势优化”