Interaction-based multi-objective quality assessment for task-specific vacuum grasping
Georg Siegemund, Florian Hecht, Franz Dietrich
Technical University Berlin, Institute of Machine Tools and Factory Management
CIRP Global Web Conference 2026 (CIRPe 2026) — Procedia CIRP · 2026

Introduction
A geometrically promising suction position is not necessarily the best position once the gripper physically interacts with the object.
Seal formation, leakage, pressure distribution, suction effort, contact force, and local compliance only become observable during the actual grasp. This work introduces an interaction-based, multi-objective framework for evaluating and re-ranking vacuum grasp candidates according to the requirements of a specific handling task.
At a glance
Geometric grasp planning and vacuum thresholds provide useful feasibility information but do not fully describe how robust, efficient, or gentle a grasp is.
Point-cloud-generated candidates are physically executed. Sensor information is transformed into several interpretable grasp-quality components and combined differently depending on the task.
Only 7.7% top-1 agreement was observed between the initial geometric ranking and the measured interaction-based ranking, showing that physical interaction provides important information that geometry alone cannot capture.
Beyond binary grasp success
Two successful grasps can behave very differently: one may require more suction power, another may create excessive force on the object, and a third may form a more uniform and robust seal.
A simple success/failure label loses this information, even though it is directly relevant to industrial process design.
A multi-objective grasp quality
Grasp quality is broken down into five interpretable components: seal strength, pressure uniformity, suction efficiency, gentle handling (low interaction force), and geometric robustness.
Each component captures a different aspect of how the grasp actually behaved during execution, rather than a single aggregate score.
Geometry proposes, interaction evaluates
The framework works in two stages. First, point-cloud geometry generates plausible grasp candidates for a given object.
Second, these candidates are physically executed and evaluated using real pressure, force, pose, gripper orientation, and suction-command data, producing a measured quality label that can re-rank the original candidate set.
Task-specific grasp selection
The best grasp depends on what the process actually values — robust transport, energy-efficient handling, gentle handling, or precision placement can each favour a different grasp position from the same candidate set.
Weighting the five quality components differently therefore allows the same measured database to serve different task requirements.
Learning from measured grasps
The resulting grasp database, built from physically measured quality labels, can be used as training data for a prediction model.
This demonstrates how measured grasp quality could eventually guide a robot toward promising grasp regions without exhaustively testing every position.
In the paper this stage is a demonstration only: a small neural network shows that the selected grasp position changes with the optimisation objective, without reporting predictor accuracy.
Highlights
- Grasp quality is split into five interpretable components — seal strength, pressure uniformity, suction effort, gentle handling, and geometric robustness — combined with a binary feasibility gate.
- Five task profiles weight the components differently: balanced, robust transport, energy-efficient, gentle handling, and precision placement.
- For the balanced profile, the geometric top candidate was also the measured best candidate in only 1 of 13 candidate sets (7.7%); the mean Spearman rank correlation between the two rankings was 0.07.
- Pre-lift vacuum agreed only partly with lifting success, while the pressure-uniformity map followed the success-rate map more closely.
Methodology
- Candidate generation: after cropping, outlier removal, and clustering, locally planar regions are detected with a plane fit inside the suction-cup footprint. A candidate needs at least 80 surface points in the cup and receives a geometric cost from planarity, distance to the object centroid, footprint coverage, and flatness penalties.
- Execution: a UR10e robot with a Formhand adaptive suction gripper executes the candidates. The system records global vacuum-chamber pressure, four local pressure signals in the gripper, force/torque, normalised suction command, and robot pose; an Ensenso S camera provides the point clouds.
- Quality labels: the recorded signals become five normalised components between 0 and 1. Task-weighted scores combine them; if the object is not retained after lifting, all scores are set to zero.
- Raster experiments: the test part was sampled on a 50 mm grid with gripper tilts from −10° to +10° and suction commands from 20% to 100%, with at least six repetitions per position and setting.
- Ranking comparison: 13 candidate sets (5 sets of 16 and 8 sets of 15 candidates), 200 executed grasps in total, comparing the geometric rank with the measured rank.
- Learning demonstration: a small neural network (8 → 64 → 32 → 1) predicts grasp feasibility from position, orientation, suction effort, and interaction force; it was trained on 1,000 sampled grasps.
Findings
- The quality landscape is spatially heterogeneous. Several regions reached high vacuum before lifting but did not lift reliably, so a vacuum threshold alone can overestimate feasibility.
- Pressure uniformity related more strongly to the measured success rate than pre-lift vacuum did.
- Changing the suction command or tilt did not simply improve the map; it shifted the regions of high and low uniformity.
- Some candidates with good geometric ranks showed leakage or low pressure uniformity, while several lower-ranked candidates achieved more stable seals and higher measured quality.
- The mismatch between geometric and measured top candidates persisted across the other task profiles.
- In the learning demonstration, the selected grasp position changed when only the optimisation objective was changed.
Figures


Method diagram

Limitations
- All quantitative results were obtained on a single 3D-printed test part; the other objects shown illustrate applicability only.
- The local pressure sensors are integrated into the gripper, which may not transfer to every suction-cup design.
- The framework should be understood as object- or object-family-specific until generalisation to unseen objects, gripper geometries, materials, and sensing configurations is validated.
- The learning stage is a demonstration; predictor accuracy is not evaluated.
Citation
Siegemund, G., Hecht, F., & Dietrich, F. (2026). Interaction-based multi-objective quality assessment for task-specific vacuum grasping. Procedia CIRP — CIRP Global Web Conference 2026 (CIRPe 2026).
Interested in a research collaboration?
Get in touch to discuss joint research, data collection, or applied evaluation of new methods.