Liangliang Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Liangliang Wang has a total of 9 co-authors. Here are the top co-authors by number of shared publications:
- Guanglei Huo
Fujian (Quanzhou) Institute of Advanced Manufacturing Technology; Fujian Key Laboratory of Intelligent Operation and Maintenance Robot Technology, Quanzhou, China
1 shared publication - Jiayao Liu
Beijing Jiaotong University, Beijing, China
1 shared publication - Jinhu Dang
Institute of Smart Manufacturing Systems, Chang’an University, Xi’an, Shaanxi 710064, People’s Republic of China
1 shared publication - Mingkun Xu
Institute of Smart Manufacturing Systems, Chang’an University, Xi’an, Shaanxi 710064, People’s Republic of China
1 shared publication - Peidong Liang
Fujian (Quanzhou) Institute of Advanced Manufacturing Technology, Quanzhou, Fujian 362000, People’s Republic of China
1 shared publication - Wanjin Guo
Institute of Smart Manufacturing Systems, Chang’an University, Xi’an, Shaanxi 710064, People’s Republic of China
1 shared publication
Latest Publications
- 8 May 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (25)
A real-time pixel-level grasp detection method for unordered stacked scenarios
In unordered stacked scenarios, object overlap and occlusion hinder accurate grasp detection, while improved detection accuracy alone cannot guarantee safe robotic grasping. Meanwhile, existing methods mainly predict discrete gripper configurations and often neglect inter-object manipulation relationships, which may result in collisions and consequent grasp failure. To address this issue, this article proposes a real-time pixel-level grasp detection method tailored for unordered stacked scenarios. First, a novel grasp representation is proposed to represent gripper configurations, enabling better observation of the mapping of the gripper’s grasp positions on the object during network prediction. Then, a grasp path labeling method is proposed to streamline pixel-level dataset labeling. Subsequently, the multi-scale feature fusion adaptive perceptual network (MFFAP-Net) is proposed, which processes each image within 13 ms. The network takes RGB images as input, extracts and processes features through a combination of hybrid attention pyramid and multi-scale fusion decoder, and outputs the grasp configurations of each pixel. Finally, the optimal grasp points are screened by an object relationship detection module. The network achieves 99.4% grasp detection accuracy on the Cornell dataset and demonstrates robustness to unseen objects and environmental variations. Grasping experiments conducted on a UR10e collaborative robot achieve an average grasp success rate of 96.81% in single-object scenarios, 94.86% in cluttered scenarios, and 73.67% in stacked scenarios.
- 8 May 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (25)
A real-time pixel-level grasp detection method for unordered stacked scenarios
In unordered stacked scenarios, object overlap and occlusion hinder accurate grasp detection, while improved detection accuracy alone cannot guarantee safe robotic grasping. Meanwhile, existing methods mainly predict discrete gripper configurations and often neglect inter-object manipulation relationships, which may result in collisions and consequent grasp failure. To address this issue, this article proposes a real-time pixel-level grasp detection method tailored for unordered stacked scenarios. First, a novel grasp representation is proposed to represent gripper configurations, enabling better observation of the mapping of the gripper’s grasp positions on the object during network prediction. Then, a grasp path labeling method is proposed to streamline pixel-level dataset labeling. Subsequently, the multi-scale feature fusion adaptive perceptual network (MFFAP-Net) is proposed, which processes each image within 13 ms. The network takes RGB images as input, extracts and processes features through a combination of hybrid attention pyramid and multi-scale fusion decoder, and outputs the grasp configurations of each pixel. Finally, the optimal grasp points are screened by an object relationship detection module. The network achieves 99.4% grasp detection accuracy on the Cornell dataset and demonstrates robustness to unseen objects and environmental variations. Grasping experiments conducted on a UR10e collaborative robot achieve an average grasp success rate of 96.81% in single-object scenarios, 94.86% in cluttered scenarios, and 73.67% in stacked scenarios.
- Open Access
- 13 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-4
GCN Rewarded Reinforcement Learning for Action Segmentation
- 13 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-4
Open AccessGCN Rewarded Reinforcement Learning for Action Segmentation