Training-free Suction Grasp Detection for Deformed Aseptic Cartons Using Vision-Language Models and Geometric Surface Scoring

Marin Maletic; Goran Vasiljevic

cs.ROAug 28, 2026

Training-free Suction Grasp Detection for Deformed Aseptic Cartons Using Vision-Language Models and Geometric Surface Scoring

Authors: Marin Maletic, Goran Vasiljevic

Organizations: Laboratory for Robotics and Intelligent Control Systems University of Zagreb Faculty of Electrical Engineering and Computing, Unska 3, 10000 Zagreb, Croatia

Abstract

Robotic sorting of recyclable waste is challenging due to the deformable and geometrically inconsistent nature of target objects. We present a training-free suction grasping system for sorting deformed aseptic beverage cartons, decoupling target identification from grasp-point selection. An open-vocabulary vision-language model detects cartons from a text prompt, SAM2 refines each detection into an instance mask, and a geometric scoring method selects the suction point by combining surface flatness with normal alignment. Three geometric methods are compared: k-nearest-neighbour PCA, Sobel cross-product, and RANSAC plane fitting. Evaluated on a real robot across three deformation levels and 35 cluttered scenes, single-object grasp success reaches 88.2% and end-to-end retrieval in clutter is 72.6%.

Figures & tables

Figure 1: Experimental setup: (a) robotic workcell (b) successful suction grasp of heavily deformed carton

Figure 2: Overview of the pipeline. Perception: The depth camera’s RGB image is passed with a prompt to the VLM and the resulting bounding boxes are forwarded to SAM2 for segmentation. Surface analysis: Depth pixels inside the mask are back-projected into a segmented point cloud, on which surface normals (red) and local flatness are estimated by one of three geometric methods, yielding a per-point suction quality heatmap. Grasp candidate selection: among points clearing the sealing threshold, the one closest to the object centroid is chosen to minimise lift torque. Grasp execution: The resulting pose is sent to the robot arm, which approaches vertically, forms a vacuum seal, and lifts the carton.

Figure 3: Examples of different carton deformation levels

Lvl	Method	$N$	SD (mm)	Time (ms)
1	KNN	20	7.0	$237\pm 50$
	Sobel	20	11.9	$\mathbf{127\pm 6}$
	RANSAC	20	4.5	$5610\pm 3157$
2	KNN	30	12.4	$266\pm 65$
	Sobel	30	17.1	$\mathbf{135\pm 7}$
	RANSAC	30	12.2	$6105\pm 3423$

Table 1: Scoring consistency and runtime per method per deformation level. Consistency is the mean 3-D standard deviation (SD) of the selected suction point across 15 repeated runs, averaged over all

N

trials (in mm). Scoring time is the per-frame computation time (in ms).

Lvl	KNN	Sobel	RANSAC
1	30/30 (1.00)	30/30 (1.00)	30/30 (1.00)
	[.89, 1.00]	[.89, 1.00]	[.89, 1.00]
2	26/30 (.87)	27/30 (.90)	26/30 (.87)
	[.70, .95]	[.74, .97]	[.70, .95]
3	23/33 (.70)	25/33 (.76)	26/33 (.79)
	[.53, .83]	[.59, .87]	[.62, .89]

Table 2: Single-object grasp success rate per scoring method per deformation level. Shown as successes/attempts (rate). Brackets give Wilson 95% confidence intervals.

Figure 4: Exp. 3 - Detection and scoring visualized

Cartons	Sc.	Det. recall	Grasp/Det.	E2E
2-4	15	86.7%	92.3%	80.0%
5-8	20	82.3%	85.0%	70.0%
175	35	83.4%	87.0%	72.6%

Table 3: Cluttered-scene performance across 35 scenes Sc. (15 sparse and 20 dense) containing 175 cartons total, mixed among distractors. Detection recall = detected/present; conditional grasp success = grasped/detected; end-to-end retrieval = grasped/present.

Figure 5: Prompt-driven target selection - detection and segmentation outputs for different natural-language prompts

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

Monocular Vision Based Control Framework for Grasping
Jul 8, 2026Shail Jadav, Dongheui LeeRobotic Grasping Deformable Object Manipulation
Autonomous Systems, Technische Universität Wien (TU Wien), Vienna, Austria · Institute of Robotics and Mechatronics (DLR), German Aerospace Center, Wessling, Germany
Beyond the Current Scene: Event-Referential Grasping with Active View Selection
Sep 30, 2026Hyunjoon Lee, Haebeom Jung, Eunsung Cha +4Robotic Grasping Language-Conditioned Robot Manipulation
Seoul National University · NVIDIA
SeededGrasp: Language-Guided Grasping in Complex Scenes with Multiple Embodiments
Jul 22, 2026Yang Xu, Gurpreet Singh Mukker, Raymond Wang +3Robotic Grasping Language-Conditioned Robot Manipulation
University of Toronto · Vector Institute · University of British Columbia +1

Jul 8, 2026J/K move · Enter open · S save

Monocular Vision Based Control Framework for Grasping

Shail Jadav, Dongheui Lee

Autonomous Systems, Technische Universität Wien (TU Wien), Vienna, Austria · Institute of Robotics and Mechatronics (DLR), German Aerospace Center, Wessling, Germany

Robotic Grasping Deformable Object Manipulation Vision-Based Robot Control

Open paper AskPDF arXiv

Grasping in unstructured environments requires handling objects with widely different mechanical properties, from soft and deformable items to rigid everyday objects. Most existing approaches address these categories separately and often rely on tactile sensing, object-specific models, or specialized grippers. In this paper, we present a unified monocular vision-based grasping framework that targets both soft and rigid objects within a single control pipeline, using only RGB input and a position-controlled gripper. The proposed system combines open-vocabulary object detection, image segmentation, boundary-aware point assignment, real-time point tracking, and monocular depth estimation to recover object motion and geometry from visual observations. A key component of the framework is a language-based stiffness estimation model that infers an object's expected compliance from its semantic description and provides an object-level prior for selecting the grasping strategy before contact. For deformable objects, grasp adaptation is governed by a Procrustes-based dissimilarity measure computed from tracked keypoints, which acts as a visual proxy for deformation. For rigid objects, the gripper width is regulated through the scaling of tracked point distances. We validate the proposed method in real-world pick-and-place experiments on a Franka Emika Research 3 arm using objects with substantially different mechanical properties, including lettuce, fresh mozzarella cheese, croissants, paper towels, and hard plastic bottles. Results demonstrate that the framework achieves stable grasping across both soft and rigid objects using visual feedback alone, highlighting a practical, sensor-efficient, and generalizable approach for food handling and household manipulation.

Abstract

Robotic Grasping

Explore similar work

Monocular Vision Based Control Framework for Grasping

Beyond the Current Scene: Event-Referential Grasping with Active View Selection

SeededGrasp: Language-Guided Grasping in Complex Scenes with Multiple Embodiments