GAPG: Geometry Aware Push-Grasping Synergy for Goal-Oriented Manipulation in Clutter

Lijingze Xiao; Jinhong Du; Yang Cong; Supeng Diao; Yu Ren

cs.ROMar 22, 2026

GAPG: Geometry Aware Push-Grasping Synergy for Goal-Oriented Manipulation in Clutter

Authors: Lijingze Xiao, Jinhong Du, Yang Cong, Supeng Diao, Yu Ren

Organizations: College of Automation Science and Engineering, South China University of Technology, Guangzhou, China.

Abstract

Grasping target objects is a fundamental skill for robotic manipulation, but in cluttered environments with stacked or occluded objects, a single-step grasp is often insufficient. To address this, previous work has introduced pushing as an auxiliary action to create graspable space. However, these methods often struggle with both stability and efficiency because they neglect the scene's geometric information, which is essential for evaluating grasp robustness and ensuring that pushing actions are safe and effective. To this end, we propose a geometry-aware push-grasp synergy framework that leverages point cloud data to integrate grasp and push evaluation. Specifically, the grasp evaluation module analyzes the geometric relationship between the gripper's point cloud and the points enclosed within its closing region to determine grasp feasibility and stability. Guided by this, the push evaluation module predicts how pushing actions influence future graspable space, enabling the robot to select actions that reliably transform non-graspable states into graspable ones. By jointly reasoning about geometry in both grasping and pushing, our framework achieves safer, more efficient, and more reliable manipulation in cluttered settings. Our method is extensively tested in simulation and real-world environments in various scenarios. Experimental results demonstrate that our model generalizes well to real-world scenes and unseen objects.

Figures & tables

Fig. 1 : This figure illustrates the GAPG workflow. The objective is to extract target objects with an asterisk label from dense scenes. Since the target object is not graspable in the initial state (State 1), we use pushing actions to adjust the object’s spatial configuration to state (State 4). During this process, both grasping and pushing actions are derived from point cloud sampling.

Fig. 2 : Overview of our framework. Grasp Evaluation Module : The model takes as input the global point cloud and the target point cloud.Based on the grasp pose, a gripper point cloud is generated and concatenated with the point cloud within the gripper’s closure space to form a grasp representation. This representation is then fed into a PointNet++ network to extract global geometric features, which are finally passed to an MLP for grasp feasibility analysis. Push Evaluation Module : The push pose is converted into a fixed push point, and the same spatial transformation is applied to the global point cloud. The transformed data is then concatenated with one-hot labels to distinguish the push point/target object from other objects. The synthesized push state is processed through PointNet++ to extract (N+1) × 192 features, from which the 1 × 192 feature corresponding to the push point is selected and fed into an MLP to score the push pose.

Fig. 3 : Geometric matching between the gripper and the target object’s grasping surface: In the left figure, the gripper fingertips are parallel to the target object’s grasping surface, with a large contact area and uniform force distribution, resulting in a high geometric match, thus deemed graspable. In the right figure, when the gripper closes, the contact area with the target object is smaller, and the fingertips struggle to form a stable grasp on the surface, leading to a low geometric match, thus deemed non-graspable.

Fig. 4 : Push action value score: In the left figure, due to the highly dense state, it is difficult to make the target object graspable through the current push, resulting in a low score . In the right figure, in a moderately dense state, the current push can enable the target object to reach a graspable state in the next step, thus resulting in a higher score.

Fig. 5 : In the PyBullet simulation environment, eight artificially created challenge scenes and four scenes loaded randomly with 15 unseen objects. The objects marked with an asterisk in these scenes are the target objects.

Panel of Figure 6 — Fig. 6 : Module training and ablation experiments: The module converged after 80 training epochs. Visually, our system outperforms the ablation methods by a significant margin, both in grasp evaluation accuracy and push evaluation accuracy.

Fig. 7 : Three types of scenes were arranged in the real-world setting, with the target objects marked by an asterisk. Orange frame: challenge scenes manually constructed; Purple frame: cluttered and dense scenes with 15 objects manually arranged; Red frame: 15 objects not seen in the training dataset and with geometric shapes significantly different from those in the training objects.

Method	Task success rate			Grasp success rate			Average motion number
Method	15 objects	30 objects	challenge	15 objects	30 objects	challenge	15 objects	30 objects	challenge
GAPG-Single Grasping	43.3%	13.3%	/	100%	100%	/	/	/	/
Xu et al. [ 8 ] -Single Grasping	26%	12%	/	85%	79.3%	/	/	/	/
Efficient Push Grasping [ 8 ]	81%	76%	86%	84%	80.1%	87%	3.93	4.93	2.93
GAPG w/o gripper pc	76.7%	60%	90%	76.7%	64.3%	90%	2.90	5.00	2.60
GAPG	100%	100%	100%	100%	96.7%	100%	1.63	2.80	2.33

TABLE I : Comparison of different methods in simulation

Panel of Figure 9 — Fig. 8 : Results of generalization experiments on unseen objects. Figure (a) shows the experimental results in a simulated environment with a scene containing 15 unseen objects; Figure (b) presents the experimental results in a real-world scenario with a physically arranged scene of 15 unseen objects.

Method	Efficient Push-Grasping	GAPG
Task success rate (15 objects)	65%	95%
Task success rate (challenge)	70%	100%
Grasp success rate (15 objects)	81.3%	100%
Grasp success rate (challenge)	82.4%	100%
Average motion number (15 objects)	5.75	2.75
Average motion number (challenge)	4.60	2.50

TABLE II : Comparison of different methods in real-world

Robotic Grasping

Robotic grasping studies how robots perceive, plan, and execute stable, task-appropriate contacts with objects across varied shapes and conditions.

See all 311 papers

AdaClearGrasp: Learning Adaptive Clearing for Zero-Shot Robust Dexterous Grasping in Densely Cluttered Environments
Mar 11, 2026Zixuan Chen, Wenquan Zhang, Jing Fang +7Robotic Grasping Robotics
Nanjing University · Nanjing University of Posts and Telecommunications · National University of Singapore
T(R,O) Grasp++: Cross-Embodiment Dexterous Grasping in Cluttered Environments at Scale
Oct 18, 2026Xin Fei, Zhixuan Xu, Huaicong Fang +5Robotic Grasping Cross-Embodiment Robot Learning
Department of Computer Science, National University of Singapore · RoboScience · College of Control Science and Engineering, Zhejiang University, China +2
AeroGrab: A Unified Framework for Aerial Grasping in Cluttered Environments
Mar 16, 2026Shivansh Pratap Singh, Naveen Sudheer Nair, Samaksh Ujjawal +4Aerial Robotics Robotic Grasping
Robotics Research Center, IIIT Hyderabad, India. · Department of Computer Science, University of Manchester, UK.

Mar 11, 2026J/K move · Enter open · S save

AdaClearGrasp: Learning Adaptive Clearing for Zero-Shot Robust Dexterous Grasping in Densely Cluttered Environments

Zixuan Chen, Wenquan Zhang, Jing Fang, Ruiming Zeng+6

Nanjing University · Nanjing University of Posts and Telecommunications · National University of Singapore

Robotic Grasping Robotics Zero-Shot Robotic Manipulation RL for Robotics Robot Manipulation Benchmarks

Open paper AskPDF arXiv

In densely cluttered environments, physical interference, visual occlusions, and unstable contacts often cause direct dexterous grasping to fail, while aggressive singulation strategies may compromise safety. Enabling robots to adaptively decide whether to clear surrounding objects or directly grasp the target is therefore crucial for robust manipulation. We propose AdaClearGrasp, a closed-loop decision-execution framework for adaptive clearing and zero-shot dexterous grasping in densely cluttered environments. The framework formulates manipulation as a controllable high-level decision process that determines whether to directly grasp the target or first clear surrounding objects. A pretrained vision-language model (VLM) interprets visual observations and language task descriptions to reason about grasp interference and generate a high-level planning skeleton, which invokes structured atomic skills through a unified action interface. For dexterous grasping, we train a reinforcement learning policy with a relative hand-object distance representation, enabling zero-shot generalization across diverse object geometries and physical properties. During execution, visual feedback monitors outcomes and triggers replanning upon failures, forming a closed-loop correction mechanism. To evaluate language-conditioned dexterous grasping in clutter, we introduce Clutter-Bench, the first simulation benchmark with graded clutter complexity. It includes seven target objects across three clutter levels, yielding 210 task scenarios. We further perform sim-to-real experiments on three objects under three clutter levels (18 scenarios). Results demonstrate that AdaClearGrasp significantly improves grasp success rates in densely cluttered environments. For more videos and code, please visit our project website: https://chenzixuan99.github.io/adaclear-grasp.github.io/.

Abstract

Robotic Grasping

Explore similar work

AdaClearGrasp: Learning Adaptive Clearing for Zero-Shot Robust Dexterous Grasping in Densely Cluttered Environments

T(R,O) Grasp++: Cross-Embodiment Dexterous Grasping in Cluttered Environments at Scale

AeroGrab: A Unified Framework for Aerial Grasping in Cluttered Environments