OPG-Policy: Occluded Push-Grasp Policy Learning with Amodal Segmentation

Hao Ding, Yiming Zeng, Zhaoliang Wan, Hui Cheng

Abstract

Robotic grasping in cluttered environments is often hindered by partial occlusions, where target objects are hidden by obstacles. Traditional methods relying on visible segments often lead to inefficient retrieval. We present OPG-Policy, a framework that utilizes amodal segmentation to predict the full shape of partially hidden targets. By integrating these amodal masks into a Deep Q-Learning framework with a dedicated action coordinator, our system achieves superior success rates and motion efficiency. Experimental results in simulation and real-world settings demonstrate that OPG-Policy significantly outperforms existing state-of-the-art push-grasp policies.

Introduction

In real-world scenarios like warehouse bin picking or household tidying, objects are rarely fully visible. For a robot, a target object might only be partially exposed, leading to ambiguity regarding its center of mass, orientation, and stable grasp points. Conventional grasping policies that operate solely on visible segments often result in suboptimal actions, such as pushing obstacles into the target or attempting unstable grasps. This "low motion efficiency" is a major bottleneck in robotic manipulation.

Methodology

The OPG-Policy framework addresses occlusion through three integrated components:

Amodal Segmentation Module

The system first predicts the amodal mask of the target object—the union of its visible and occluded parts. This provides the robot with a spatial prior of the object's true extent. Training data for this module is generated in simulation using the Segment Anything Model (SAM) to label visible parts and the ground truth for hidden parts.

Deep Q-Network (DQN) with Amodal Input

The policy takes three heightmaps (RGB, Depth, and Amodal Mask) as input. A DenseNet backbone extracts features to generate Q-maps for two primitive actions: Pushing ($Q_p$) and Grasping ($Q_g$). Each pixel in these maps represents the expected reward for an action at a specific location and orientation.

The OPG-Policy architecture integrates amodal segmentation with a dual-stream DQN and a coordinator MLP to select optimal actions.

Action Coordinator

A three-layer Multi-Layer Perceptron (MLP) acts as a coordinator. It evaluates the best Q-values alongside domain-specific features: the occlusion rate and the occlusion border. This allows the robot to strategically decide whether to continue clearing obstacles (pushing) or attempt a retrieval (grasping).

Experimental Results

The OPG-Policy was evaluated against baselines such as GTI and GE-GRASP in various cluttered scenarios.

Performance and Efficiency

As shown in the comparative analysis, OPG-Policy maintains high success rates even as occlusion increases. While other methods see a sharp decline in performance under heavy occlusion (0.6-0.8 ratio), OPG-Policy remains robust. Furthermore, it achieves these results with fewer total actions, indicating higher motion efficiency.

Comparative results showing success rates and total attempts across different occlusion ratios for OPG-Policy and baseline methods.

Real-World Generalization

The model, trained entirely in simulation, was transferred to a real-world UR10 robot setup without fine-tuning. In tests with unseen YCB objects, it achieved a 90% success rate with an average of 6.65 moves per retrieval, demonstrating strong zero-shot transfer capabilities.

Conclusion

OPG-Policy demonstrates that "seeing through" clutter via amodal segmentation significantly enhances robotic grasping efficiency. By combining amodal priors with a strategic action coordinator, robots can navigate dense clutter more effectively than policies relying on partial observations. Future work will explore extending this framework to multi-object sorting and handling cases of total occlusion.

Summary updated

BibTeX
@inproceedings{ding2024opg,
  author = {Hao Ding and Yiming Zeng and Zhaoliang Wan and Hui Cheng},
  title = {{OPG-Policy: Occluded Push-Grasp Policy Learning with Amodal Segmentation}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  url = {https://openreview.net/forum?id=allHIuP4zm},
}
Figure