A Real-World Grasping-in-Clutter Performance Evaluation Benchmark for Robotic Food Waste Sorting

Moniesha Thilakarathna; Xing Wang; Min Wang; David Hinwood; Shuangzhe Liu; Damith Herath

cs.ROFeb 21, 2026

A Real-World Grasping-in-Clutter Performance Evaluation Benchmark for Robotic Food Waste Sorting

Authors: Moniesha Thilakarathna, Xing Wang, Min Wang, David Hinwood, Shuangzhe Liu, Damith Herath

Organizations: Faculty of Science and Technology, University of Canberra, 11 Kirinari St, Bruce ACT 2617, Australia

Abstract

Food waste management is critical for sustainability, yet inorganic contaminants hinder recycling potential. Robotic automation accelerates sorting through automated contaminant removal. Nevertheless, the diverse and unpredictable nature of contaminants introduces major challenges for reliable robotic grasping. Grasp performance benchmarking provides a rigorous methodology for evaluating these challenges in underexplored field contexts like food waste sorting. However, existing approaches suffer from limited simulation datasets, over-reliance on simplistic metrics like success rate, inability to account for object-related pre-grasp conditions, and lack of comprehensive failure analysis. To address these gaps, this work introduces GRAB, a real-world grasping-in-clutter (GIC) performance benchmark incorporating: (1) diverse deformable object datasets, (2) advanced 6D grasp pose estimation, and (3) explicit evaluation of pre-grasp conditions through graspability metrics. The benchmark compares industrial grasping across three gripper modalities through 1,750 grasp attempts across four randomized clutter levels. Results reveal a clear hierarchy among graspability parameters, with object quality emerging as the dominant factor governing grasp performance across modalities. Failure mode analysis shows that physical interaction constraints, rather than perception or control limitations, constitute the primary source of grasp failures in cluttered environments. By enabling identification of dominant factors influencing grasp performance, GRAB provides a principled foundation for designing robust, adaptive grasping systems for complex, cluttered food waste sorting.

Figures & tables

Figure 1: Industrial-scale variation of inorganic contaminants inside food waste

Panel of Table 1 — Table 1: Comparison of grasp performance benchmarking studies.

Figure 2: Overview of the proposed GRAB benchmark for grasping-in-clutter, illustrating the integration of (A) a real-world object dataset spanning seven categories, (B) an experimental robotic grasping setup with RGB-D perception and interchangeable grippers, (C) grasping modalities (rigid, soft adaptive, and suction), and (D) graspability parameters ( $Q_{o}$ , $Q_{p}$ , $Q_{c}$ ) defining pre-grasp conditions. Performance is evaluated using success ( $S_{P}$ ), stability ( $S_{B}$ ), and efficiency ( $S_{F}$ ) across multiple clutter levels (L1–L4), followed by failure mode analysis.

Figure 3: Conceptualization of the real-world robotic grasping pipeline implemented using the ROS 2 control stack: (a) parallel grasping path for the rigid force-controlled parallel gripper (RG2) and adaptive Fin-Ray gripper, and (b) suction grasping path for the suction gripper. The pipeline illustrates RGB-D sensor data acquisition, 6D grasp pose detection (grasp generation using AnyGrasp and SuctionNet), post-processing, and selection of top-ranked feasible grasp poses, followed by path planning and grasp execution. The MoveIt 2 block illustrates the motion planning process, including the home pose, pre-grasp pose (purple arm), grasp pose within the waste scene, post-grasp pose, and release pose.

Figure 4: Three distinct grasping modes (a) Rigid force-controlled two-finger parallel gripper (RG2) with FSR UX 400 force sensor (b) Soft adaptive two-finger Fin-Ray gripper with FSR UX 408 force sensor (c) Suction gripper with Honeywell APB pressure sensor

Figure 5: Key stages of the grasp execution process: (1) home pose with grasp pose visualization in RViz, (2) pre-grasp pose, (3) grasp pose during object engagement, (4) post-grasp pose after successful object lifting, and (5) release pose.

Figure 6: Overview of 3D scanning and point-cloud processing workflow for DCD computation (a) Customized 3D scanning and compression platform, (b) RecFusion 3D reconstruction, (c) point cloud of the undeformed object before gripping, (d) point cloud of the deformed object under compression, and (e) aligned point clouds illustrating geometric differences due to deformation.

Figure 7: Object dataset with their corresponding DCD values

Figure 8: Variation of occupancy during a grasp cycle for an object cluster

Figure 9: Object-holding tracking system for grasp stability evaluation, showing force sensor measurements during grasp execution: (a) successful transfer, where the object is retained from grasp pose ( $t_{1}$ ) to release pose ( $t_{3}$ ) with $t_{2}=t_{3}$ , and (b) mid-way drop, where the object is released prematurely due to slippage ( $t_{2}<t_{3}$ )

Panel of Table 2 — Table 2: Evaluation of grasp performance using radar charts. Performance polygons are color-coded by gripper (blue: rigid, pink: Fin-Ray, green: suction), where the gripper with the largest enclosed area represents the highest overall performance, indicated by the corresponding background color.

Figure 10: Exploratory data analysis of graspability parameters. (a) Kernel density distributions of object score $Q_{o}$ , pose score $Q_{p}$ and clutter score $Q_{c}$ (b) Correlation heatmap

	$S_{p}$			$S_{b}$			$S_{f}$
Term	coef	OR	p	coef	OR	p	coef	OR	p
$Q_{p}$	1.4543	4.2815	0.1656	1.3233	3.7557	0.1909	0.3579	1.4304	0.6345
$Q_{c}$	1.1525	3.1661	0.0790	1.1121	3.0408	0.0897	0.6832	1.9802	0.0400
$Q_{o}$	4.0842	59.3952	0.0008	4.0813	59.2247	0.0004	2.0424	7.7091	0.0101
Finray vs Rigid	1.6127	5.0165	0.0341	1.4474	4.2522	0.0331	1.1662	3.2098	0.0293
$Q_{p}\times$ Finray	-0.4471	0.6395	0.6879	-0.3319	0.7175	0.7465	0.1092	1.1154	0.8907

Table 3: Regression results for grasp performance metrics. Grasp success (

S_{p}

) is modeled using logistic regression, while stability (

S_{b}

) and efficiency (

S_{f}

) use fractional logit models. The Rigid gripper is the reference category. Green = significant positive (

p<0.05

), Red = significant negative (

p<0.05

), White = not significant.

Figure 11: Partial dependence plots illustrating the effect of graspability parameters (Object Score $Q_{o}$ , Grasp Score $Q_{p}$ , and Clutter Score $Q_{c}$ ) on the predicted probabilities of Grasp Success $S_{p}$ , Grasp Stability $S_{b}$ and Grasp Efficiency $S_{f}$ across gripper modalities (Rigid, Fin-ray, and Suction).

Figure 12: (a) Variation in three major categories of failure modes: physical interaction failures, perception failures, and execution failures observed across the three grasping modalities. (b) Variation of physical interaction failure modes: CL - gripper collisions with the work base or target object, CLS - collisions with surrounding objects, SL - slippage, and ON - failures related to object non-conformance due to non-planar or highly irregular surfaces

Figure 13: Impact of graspability parameters (pre-grasp conditions) $Q_{o}$ , $Q_{p}$ and $Q_{c}$ on the occurrence of grasp failures for three gripper modalities: (a) Rigid two-finger parallel gripper (RG2) (b) Soft adaptive two-finger Fin-Ray gripper (c) Suction gripper.

	Failure Mode	$Q_{o}$ (Object Score)	$Q_{p}$ (Pose Score)	$Q_{c}$ (Clutter Score)
Rigid / Fin-Ray	CL	$\checkmark$	$\triangle$	$\times$
	CLS	$\checkmark$	$\triangle$	$\checkmark$
	SL	$\triangle$	$\triangle$	$\triangle$
	WGP	$\checkmark$	$\triangle$	$\times$
Suction	WGP	$\checkmark$	$\triangle$	$\times$
Suction	ON	$\checkmark$	$\times$	$\times$

Table 4: Impact of object-related (Object Score

Q_{o}

), pose quality-related (Pose Score

Q_{p}

), and clutter-related (Clutter Score

Q_{c}

) pre-grasp conditions on the occurrence of failure modes across gripper types (corresponding to Fig. 13 ).

Robotic Grasping

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

See all 309 papers

Beyond Visual Grasping: Benchmarking Complex Grasping from Detection to Execution
Jul 15, 2026Hanyi Zhang, Khang Nguyen, Charith Munasinghe +10Robotic Grasping Robot Foundation Models
University of Liverpool, UK · Mohamed bin Zayed University of Artificial Intelligence, UAE · Zurich University of Applied Science, Switzerland +3
Towards Reliable Sequential Object Picking in Clutter: The Runner-up Solution to RGMC 2025
Jun 11, 2026Wei Yu, Xidan Zhang, Ziyi Zheng +2Robotic Grasping Robotics
School of Mechanical Engineering, Zhejiang University, Hangzhou, China.
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

Jul 15, 2026J/K move · Enter open · S save

Beyond Visual Grasping: Benchmarking Complex Grasping from Detection to Execution

Hanyi Zhang, Khang Nguyen, Charith Munasinghe, Basu Hela+9

University of Liverpool, UK · Mohamed bin Zayed University of Artificial Intelligence, UAE · Zurich University of Applied Science, Switzerland+3

Robotic Grasping Robot Foundation Models Robot Manipulation Benchmarks

Open paper AskPDF arXiv

Robust robotic grasping remains a fundamental challenge for complex real-world applications. Recent advances in large-scale models demonstrate promising capabilities for reasoning in robotic tasks. However, existing benchmarks for grasping primarily focus on isolated, visual-based grasp pose detection, failing to capture the complexity of grasping tasks that require multi-step reasoning and semantic understanding during execution. To address this gap, we propose GCA-Bench, a benchmark featuring challenging grasping with complex action scenarios that involve both scene-level reasoning and semantic constraints. GCA-Bench enables the evaluation of recent large foundation models under the same settings. To demonstrate the effectiveness of our new benchmark, we implement a diverse set of baselines, ranging from traditional grasp detection pipelines to end-to-end learning methods. Empirical studies achieve success rates below 70% on complex grasping scenarios, underscoring critical limitations. In addition, we propose new evaluation metrics, analyze critical failure models, and provide insights to guide the development of more robust and generalizable grasping strategies.

Abstract

Robotic Grasping

Explore similar work

Beyond Visual Grasping: Benchmarking Complex Grasping from Detection to Execution

Towards Reliable Sequential Object Picking in Clutter: The Runner-up Solution to RGMC 2025

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