Research Analyzer
← Back ICRA 2026

Learning Push-Grasp Synergy for Occluded Objects in Cluttered Environments

Ziang Li, Haorui Wu, Zhiqi Chen, Haozhe Zhang, Yuzhe Huang, Changshui Zhang

PDF

AI summary

Key figure (auto-extracted from paper)
A self-supervised reinforcement learning framework with dynamic target switching and mask-guided action selection enables robust push-grasp synergy for retrieving occluded objects in clutter.
push-grasp synergy occluded grasping deep reinforcement learning cluttered environments robotic manipulation target switching

Problem

Robots struggle to grasp targets in highly cluttered environments where severe occlusion blocks direct access and traditional push-grasp strategies become inefficient or ineffective.

Approach

The method uses a deep Q-network to learn coordinated pushing and grasping, dynamically switching to alternative targets when the goal is hidden and constraining actions with object masks to improve efficiency.

Key results

  • Dynamic target-switching mechanism for severe occlusion
  • Mask-constrained action selection reducing exploration space
  • High task completion rates under severe and complete occlusions in simulation
  • Zero-shot transfer to physical robots without retraining

Why it matters

Advances autonomous manipulation in real-world cluttered settings, benefiting logistics, warehousing, and domestic robotics.

Abstract

Successfully executing grasping tasks within highly cluttered spaces is still a significant hurdle in robotics, especially in scenarios involving severe target occlusion. To tackle this, we present a novel self-supervised framework driven by deep reinforcement learning that enables robots to acquire push–grasp synergy for reliable manipulation under occlusions. The core contribution of this research is the target switching mechanism that dynamically selects alternative targets when the goal object is severely occluded. Furthermore, we utilize a strategy for selecting actions based on object masks to reduce the action space, thereby improving efficiency and minimizing ineffective operations. Comprehensive evaluations across both simulated and physical environments confirm that our method achieves robust grasping performance under severe or complete occlusions. Notably, the learned policy is readily transferable to physical environments and generalizes effectively to previ- ously unseen objects. To guarantee experimental reproducibility and encourage further studies, our source code is available at https://github.com/lzalza/Learning-Push-Grasp-Synergy-for- Occluded-Objects-in-Cluttered-Environments and demonstra- tion videos can be viewed at https://youtu.be/hDm-vIlaymw.

Index terms

Deep Learning in Grasping and Manipulation

Related papers