Skip to content
DIS Robotics
ROBOTICS · ENGINEERING · ORIGINAL SOURCES
WORLD B-1 · TECHNICAL BRIEF
Technical BriefLab5 min read

The suction cup has not read your AI benchmark

RADAR80

Reliable picking begins at the contact patch. Suction physics and Berkeley’s Dex-Net research explain why recognizing an object is only part of grasping it.

Published
Updated
Primary Domain
Industrial & Manufacturing
Deployment Maturity
Lab
Sources
4 listed · 4 live links
Reading Time
5 min read
ManipulationComputer VisionRobot Learning

DIS points readers toward original work. Source facts, DIS interpretation, and the editorial verdict are kept distinct below.

What Changed

Source-grounded summary

A robot can identify the correct object, move to the intended location and still fail to pick it up. The contact patch gets the final vote. For suction, a plausible-looking point on a depth image has to become a usable seal on a real surface.

This explainer pairs suction-cup engineering guidance with Berkeley's Dex-Net 3.0 research, first submitted in 2017 and revised in 2018. The researchers modeled compliant suction contact and robustness to uncertainty, then used synthetic examples to train a grasp-quality model. It is a historical research case, not a newly released general-purpose picking system. [1–4]

Why It Matters

DIS interpretation

Picking is an interaction among perception, geometry, material behavior and motion. “The object is there” and “this grasp will survive the intended movement” are different propositions. A robot needs enough information to connect them.

That is why the mechanical end effector should be part of a learning system's definition. A grasp score is meaningful for a particular contact model and setup. Changing the cup, camera geometry or workpiece can change the physical problem even when the software still returns a beautifully confident number. The box is under no contractual obligation to behave like the training set.

How It Works

DIS interpretation

For a basic static suction estimate, force is pressure difference multiplied by effective sealed area: F = Δp × A. Schmalz explicitly calls this a theoretical force without safety factors. [1] It is a starting model, not the entire grasp.

  1. Establish an effective seal

    Schmalz's design guidance identifies workpiece surface, cup geometry, material and deformation as relevant to practical holding behavior. Porous parts need suction trials; friction values cannot simply be assigned universally. [2]

    Consider an illustrative DIS calculation: a circular effective area 20 mm in diameter and a pressure difference of 60 kPa give A = π × 0.01² m² and F ≈ 18.85 N. That assumes the full effective area is sealed at the stated pressure. It does not specify a rated payload. Leakage, acceleration, orientation and the applicable safety margin have not been included.

  2. Ask what tries to peel the object away

    A grasp has to resist the relevant forces and moments, not just a straight pull through an ideal cup centre. An off-centre contact can make gravity apply a turning moment. Accelerating the object changes the load again.

    Dex-Net 3.0's authors therefore evaluated both seal formation and resistance to an external wrench, meaning a combined force-and-moment load. Their robustness analysis included perturbations in pose, material properties and applied loads. [3] The useful conceptual move is to evaluate the contact under uncertainty, rather than treating a nominally flat patch as automatic success.

  3. Train against the physical question you actually care about

    The paper describes a dataset built from 1,500 3D object models, containing 2.8 million point-cloud, suction-grasp and robustness-label examples. A Grasp Quality Convolutional Neural Network learned to score candidate suction targets. [3]

    Synthetic training helps explore many candidate contacts, but its labels still inherit the assumptions of the generating model. The practical question becomes whether that model captures the surfaces, cups, disturbances and sensing conditions in the real task. A large training set expands coverage within its construction; its size does not erase an omitted failure mechanism.

  4. Keep the setup attached to the result

    Berkeley's GQ-CNN documentation warns that pretrained models are sensitive to gripper geometry, camera intrinsics and camera-to-workspace distance. It does not guarantee the same performance on other physical setups. [4]

    That is useful deployment guidance. A convincing evaluation would identify the cup, object set, perception setup, motion and failure definition, then report outcomes under those conditions. It would also distinguish reliable picks from a policy that simply refuses difficult ones. A confidence threshold can change which attempts happen; the success percentage needs that context.

The Mechanical Lens

DIS interpretation · source-supported dimensions only

Signature DIS section — only engineering dimensions available in the underlying material are shown.

Mechanism

The suction seal is a compliant physical contact. Effective sealed area, surface condition and cup deformation affect how the simple pressure-area model applies. [1,2]

Materials

Surface and cup properties affect sealing and friction. Supplier guidance calls for application-specific trials, especially with porous material. [2]

Loads

Evaluate moments and motion-dependent loads as well as nominal static lifting force. Dex-Net 3.0 explicitly includes an external-wrench model. [3]

Sensors

A learned grasp policy depends on the geometric information its camera and preprocessing provide. The project documentation identifies camera setup as a transfer constraint. [4]

Numbers That Matter

Source-supported metrics

Dex-Net 3.0 synthetic examples

2.8 million

Point-cloud, suction-grasp and robustness-label examples from the paper; not 2.8 million physical robot trials. [3]

Underlying object models

1,500 3D models

Dataset-generation scope reported in the cited paper. [3]

Deployment

DIS interpretation

  • Suction handling is an established industrial technique; the cited Dex-Net 3.0 result is a specific laboratory research system.
  • Any transfer to a new camera, cup, material set or motion profile needs evaluation of that actual setup. [2,4]

The Catch

DIS interpretation

LIMITATIONS & OPEN QUESTIONS

The paper's model assumes quasi-static mechanics, rigid non-porous objects, an isolated object on a planar work surface, and known camera and cup geometry. [3] Those assumptions matter for any proposed transfer to flexible packaging, porous material or a cluttered scene. Dataset size describes training coverage, not performance on arbitrary objects.

The pressure-area calculation is equally limited: it assumes a sealed static contact and chosen dimensions. It is not a cup selection, lifting certification or prototype test. Schmalz's application guidance includes additional factors and practical trials. [2] The case study supports a connection between physical modeling and learned grasp selection; it does not establish that either can dispense with the other.

HIT THE QUESTION BLOCK BEFORE YOU BELIEVE THE HYPE.

DIS Verdict

Editorial verdict — not source fact

DIS EDITORIAL JUDGMENT

The useful breakthrough in a picking system is often the question it learns to answer. “Where is the object?” is necessary. “Will this contact hold through the intended task?” is closer to the job. Dex-Net's contact-centered framing remains worth studying because it puts mechanics inside the learning problem. The suction cup will continue marking the exam.

Sources

Original material

Follow the original material to check methods, test conditions, and limitations. Source claims are distinct from DIS interpretation.

  • Technical Docs · J. Schmalz GmbH

    [1] Holding force

    Checked 16 November 2026. Defines theoretical F = pressure difference × effective suction area, without safety factors. DIS’s 20 mm example is illustrative.

  • Technical Docs · J. Schmalz GmbH

    [2] Design of the Suction Cup

    Checked 16 November 2026. Surface/material effects, friction, porous-part trials and limits of theoretical holding-force calculations.

  • Preprint · Mahler et al. / Berkeley AUTOLAB

    [3] Dex-Net 3.0: Computing Robust Robot Vacuum Suction Grasp Targets in Point Clouds using a New Analytic Model and Deep Learning

    arXiv:1709.06670v2

    Checked 16 November 2026. Author manuscript, v2, 13 April 2018; arXiv marks acceptance to ICRA 2018. Dataset counts and contact-model scope; no universal deployment percentage inferred.

  • Technical Docs · Berkeley AUTOLAB

    [4] GQ-CNN tutorial overview

    Checked 16 November 2026. Developer warning about sensitivity to gripper geometry, camera intrinsics and workspace distance. This is setup-specific guidance, not a new benchmark.

Primary attribution label: J. Schmalz GmbH