PaperlayerView exact revision on arXiv

Source-linked AI summary

GelSlim3.0: High-Resolution Measurement of Shape, Force and Slip in a Compact Tactile-Sensing Finger

Ian Taylor, Siyuan Dong, Alberto Rodriguez

arXiv:2103.12269v1cs.RO

TL;DR

Robotic tactile fingers need compact, high-resolution contact sensing for cluttered grasping while balancing optical and manufacturing constraints. GelSlim 3.0 addresses this by optimizing a photometric-stereo-compatible optical path and integrating analytic shape, force, and slip sensing into a modular finger. The resulting design provides multiple tactile modalities in a compact, replaceable, open-source platform, while compactness leaves residual interreflection noise and limits fine texture reconstruction.

  • Problem

    Compact tactile fingers must balance photometric-stereo optical constraints, high-resolution contact sensing, and accessible fabrication for cluttered robotic grasping.

  • Method

    The paper optimizes illumination and light-shaping geometry through optical simulation, then integrates the configuration with modular mechanical design and real-time analytic measurements.

  • Results

    GelSlim 3.0 provides compact sensing of 3D shape, incipient slip, and dense 3D contact-force distribution while retaining modalities used by earlier GelSight and GelSlim sensors.

  • Takeaways & Limitations

    The open-source, replaceable design offers a compact tactile finger for cluttered bin-picking scenarios with multiple real-time sensing modalities.

  • Takeaways & Limitations

    Compactness leaves residual interreflection noise and means reconstructable texture features must be relatively large compared with previous GelSight versions.

Abstract

from arXiv · show

This work presents a new version of the tactile-sensing finger GelSlim 3.0, which integrates the ability to sense high-resolution shape, force, and slip in a compact form factor for use with small parallel jaw grippers in cluttered bin-picking scenarios. The novel design incorporates the capability to use real-time analytic methods to measure shape, estimate the contact 3D force distribution, and detect incipient slip. To achieve a compact integration, we optimize the optical path from illumination source to camera and other geometric variables in a optical simulation environment. In particular, we optimize the illumination sources and a light shaping lens around the constraints imposed by the photometric stereo algorithm used for depth reconstruction. The optimized optical configuration is integrated into a finger design composed of robust and easily replaceable snap-to-fit fingetip module that allow for ease of manufacture, assembly, use, and repair. To stimulate future research in tactile-sensing and provide the robotics community access to reliable and easily-reproducible tactile finger with a diversity of sensing modalities, we open-source the design and software at https://github.com/mcubelab/gelslim.

I. INTRODUCTION

GelSlim 3.0 is a compact tactile-sensing finger for cluttered grasping that combines high-resolution shape, force, and slip measurement with accessible assembly and open-source design.

  • Core contribution: GelSlim 3.0 targets cluttered-environment grasping while providing real-time, high-spatial-resolution measurements of shape, force, and slip.The design is intended to be easy to fabricate, use, and maintain.
  • Core contribution: The optical path, illumination sources, and light-shaping lens are optimized around photometric-stereo constraints while retaining depth reconstruction.The optical design is evaluated against current tactile sensor designs and its trade-offs and limitations are discussed.
  • Core contribution: The finger uses accessible manufacturing methods and a small number of components assembled through a simple, modular snap-to-fit design.This supports fabrication, assembly, use, and repair of the tactile finger.
  • Core contribution: Real-time analytical methods support 3D geometry reconstruction, incipient-slip detection, and estimation of the spatial distribution of 3D contact forces.These capabilities provide multiple contact measurements within one tactile-sensing finger.
  • Core contribution: The design and software are open-sourced to address manufacturing, distribution, and use bottlenecks in tactile sensing and encourage broader robotics research.The paper presents the sensing capabilities, comparisons with other designs, limitations, and repository in subsequent sections.

II. RELATED WORK

Related work spans analytical and learning-based vision-based tactile sensors, including GelSight-derived designs and systems that provide rich contact-state data for robotic manipulation.

  • Comparative designs: GelSlim 3.0 is positioned among prior integrated sensor designs including GelSight, GelSlim 2.0, and GelSlim MPalm.Figure 2 presents models of these previous and current designs from left to right.
  • Vision-based tactile sensing: Vision-based tactile sensors combine compact instrumentation with high-spatial-resolution contact-state feedback and can work synergistically with deep learning.The literature includes compact and rugged sensor designs.
  • GelSight sensors: GelSight uses a camera to image deformation of an opaque-skinned elastomer illuminated by colored LEDs from different directions.Colored shading supports direct measurements of contact position, shape, and 3D geometry.
  • GelSight sensors: GelSlim and related GelSight developments altered the optical path using a waveguide and mirror to create a more compact wedge-shaped fingertip for cluttered picking.These designs retain rich tactile data used in manipulation tasks.
  • Other vision-based tactile sensors: Other vision-based tactile systems rely more heavily on machine-learning interpretation, including neural models for finger-position and tactile-state estimation.Examples include a soft continuum finger using a convolutional neural network and OmniTact paired with a ResNet.

III. DESIGN GOALS

The design goals address optical constraints, manufacturing difficulty, and the need for compact tactile feedback in cluttered grasping. The proposed approach combines photometric-stereo-compatible optical design with modular assembly and analytic sensing.

  • Design constraints: Photometric stereo strongly constrains tactile-sensor optical design, including illumination arrangement, field of view, and sensing area for a given thickness.These constraints can make slim robotic-finger integration difficult, although photometric stereo remains useful for dexterous manipulation.
  • Manufacturing: Previous GelSight and GelSlim designs were difficult to manufacture and assemble at scale or by inexperienced users because of complex hand fabrication.Examples include acrylic waveguide bending and custom hand-soldered illumination.
  • Compactness: Compactness should allow the finger to enter clutter and place objects precisely by squeezing between objects or separating them from the environment.This is one of the stated design goals for the proposed tactile finger.
  • Analytic measurement: Analytic measurement methods provide real-time raw images, shape, force, and slip feedback without data-expensive calibration or learning procedures.These measurements are intended to support dexterous manipulation and control.
  • Design for assembly: Design-for-assembly practices aim to make fabrication and use accessible through robust, modular design, while the optical workflow uses raytracing and optimization software.The optical process evaluates design-space constraints, simulates an initial path, and optimizes illumination-shaping parameters.

1) Photometric Stereo:

GelSlim 3.0 uses model-based photometric stereo to reconstruct tactile surface geometry, requiring controlled projection, reflectance, illumination, and albedo assumptions. Its optical design adapts these constraints to a compact finger using a wide-angle camera, a planar gel, and shaped illumination.

  • Photometric Stereo: Photometric stereo analytically reconstructs elastomer deformation in 3D from images under three stated sensing assumptions.The assumptions concern orthographic projection, shading determined by surface normals without cast shadows or interreflections, and constant surface albedo.
  • Photometric Stereo: Inhomogeneous illumination reduces depth accuracy by skewing gradients or reducing valid measurements per pixel.Intensity variance and missing illumination can make the reflectance-intensity mapping non-invertible and produce incorrect depth values.
  • Initial Conditions: A 160° Raspberry Pi wide-angle camera provides a large field of view and close focus, but introduces spherical distortion that requires preprocessing.The camera cannot completely satisfy the orthographic assumption; the distortion is corrected after image capture.
  • Initial Conditions: The hexagonal prism with angled loft provides radially symmetric illumination for three color channels, supporting the photometric stereo reflectance constraints.The initial lens geometry is scaled to maximize field of view and sensing area while respecting thickness constraints.
  • Initial Conditions: A planar sensing surface improves grasping of objects lying flush against tables or bin bottoms, while Lambertian elastomer reflectance improves surface-normal measurements.Lambertian reflectance also limits the use of total internal reflection, so direct illumination and a shaping lens are used instead.

3) Simulation and Parameter Selection:

The optical system is simulated with material, receiver, and source models, then parameterized around photometric-stereo-compatible illumination objectives. The design emphasizes homogeneous, centered, white-balanced illumination while constraining lens thickness and geometry.

  • Simulation and Parameter Selection: A receiver on the sensing surface collects simulated intensity and color distributions for evaluating the optical design.Material profiles provide refractive indices and reflectance, while manufacturer source models represent package geometry, spectral data, and source distributions.
  • Simulation and Parameter Selection: The hybrid swept spline lens produces more homogeneous illumination than flat-sided lenses while generating fewer interreflective artifacts than planar or spherical alternatives.Proof-of-concept testing also identifies a 60° viewing-angle PLCC-4 OSRAM TOPLED Black T66 LED as diffuse yet shapeable.
  • Simulation and Parameter Selection: The shaping feature’s maximum thickness is constrained to 2 mm to prevent increasing the finger’s total thickness.The optimization variables include LED position and shaping-lens parameters, including position, angle, and geometric dimensions.
  • Simulation and Parameter Selection: The optimization targets a white-point chromaticity mean, zero illuminance standard deviation, and a centered illumination centroid.These objectives encourage evenly mixed, constant-intensity illumination across the sensing surface.

4) Optimization Results:

Optimization improves illumination distribution sufficiently for low-noise photometric-stereo reconstructions near the origin, although edge illumination remains imperfect. The electrical design independently balances color-channel illumination and total intensity.

  • Optimization Results: The optimized shaping features distribute illumination more effectively across the sensor surface than the initial geometry.The full optimized distribution is not perfectly radially symmetric or homogeneous, with large edge deviations remaining.
  • Optimization Results: Low-noise 3D reconstructions are demonstrated at the origin coordinate, where the optimized design satisfies the stated first- and second-order assumptions.The result supports use of the photometric stereo algorithm within the optimized optical configuration.
  • Electrical Design: The electrical assembly uses separate 3x2 parallel LED arrays for each color channel to homogenize illumination and minimize sequential intensity loss.R1, R2, and R3 normalize relative channel illumination, while R0 modulates total intensity.

C. Mechanical Design

The mechanical design packages the optical path in a compact, grasping-oriented finger assembled from modular components. Image preprocessing corrects distortion from the shaping lens and wide-angle camera before tactile measurements are used.

  • C. Mechanical Design: The fingertip, finger-body, and finger-back encapsulate the optical path while targeting a compact form factor.A curved distal fingertip helps the finger slip between objects in clutter, and the mounting geometry can adapt to various grippers.
  • C. Mechanical Design: The ten-component assembly combines 3D-printed, made-to-order, and off-the-shelf parts in a modular snap-to-fit design.Embedding heat inserts in the fingertip enables rapid assembly; the elastomer is the exception to the listed non-elastomer component sourcing categories.
  • C. Mechanical Design: The elastomer is cast from platinum-based silicone, coated with a reflective silicone layer, and fitted with laser-cut tracking markers.The marker layer is sealed with an additional silicone-ink coating.
  • A. Pre-Processing: Preprocessing corrects optical distortion by detecting the known marker array and applying a transformation matrix to the raw tactile image.The distortion originates from both the acrylic shaping lens and the wide-angle camera.

B. Shape and Geometry

GelSlim 3.0 reconstructs local 3D geometry from tactile images and supports real-time shear estimation and incipient-slip detection through marker-motion analysis.

  • Shape and Geometry: GelSlim 3.0 reconstructs local 3D geometry from RGB tactile images for contacted objects including a ball bearing, screw head, and screwdriver handle.The reconstructed geometry is presented alongside raw tactile imprints.
  • Shear and Incipient Slip: Approximate shear force is estimated by tracking the motion of the tactile sensor’s marker array.
  • Shear and Incipient Slip: During stiction, the object and sensing surface move together as a rigid body; impending slip causes part of the contact patch to lose stiction.
  • Shear and Incipient Slip: Incipient slip is detected when measured marker motion significantly deviates from motion predicted by rigid-body constraints.The method compares real contact-region motion with an estimated rigid-body motion field.

D. Dense Force Distribution

GelSlim 3.0 estimates dense 3D contact-force distributions by combining marker-derived gel motion, depth deformation, and a finite-element model, while retaining a compact design.

  • Dense Force Distribution: Dense force fields are estimated by pairing the gel’s 2D marker motion and z-axis deformation with a finite-element model of the gel.The approach augments the force-estimation method used in GelSlim 2.0 with GelSlim 3.0 depth reconstruction.
  • Dense Force Distribution: The finite-element model discretizes the gel into 8-node hexahedron elements and derives stiffness from measured Young’s modulus and Poisson’s ratio.Top-node lateral displacements come from marker tracking, while δz comes from the depth map.
  • Dense Force Distribution: F = KU maps the directly observable displacement matrix U to the force field through the stiffness matrix K.Here, U contains the displacements of all model nodes.
  • Design Comparison, Limitations and Trade-offs: GelSlim 3.0 provides twice Digit’s sensing area at a comparable thickness and price point while incorporating high-resolution images, depth, and force.
  • Design Comparison, Limitations and Trade-offs: Compactness leaves residual interreflection noise in reconstruction, limiting reliable texture capture to relatively large features compared with earlier GelSight versions.The authors identify this as a constraint on further thickness reduction using photometric stereo.

VII. OPEN-SOURCE TACTILE-SENSING FINGER

The authors open-source the GelSlim 3.0 design, production files, documentation, and software for fabricating and using the tactile finger. The sensor combines depth reconstruction, dense force distribution estimation, and incipient slip detection.

  • Open-source resources: The GelSlim repository provides documentation, design schematics, production files, a GUI, and a Raspberry Pi data-collection tutorial.These resources cover fabrication, sensor interfacing, and setup.
  • Design and deployment: The compact finger uses replaceable modules and design-for-assembly methods suited to cluttered bin-picking scenarios.The design is intended to be easy to manufacture, use, and maintain.
  • Integrated sensing: The sensor supports 3D depth reconstruction, incipient slip and shear detection, and dense force distribution estimation.These are the three integrated tactile-sensing algorithms made accessible through the software.

APPENDIX

The appendix directs readers to the repository folder containing the simulation and optimization details and resulting configuration values.

  • Simulation settings are provided in the repository’s Simulation and Optimization folder.
  • The folder also contains the optimization parameters used in the work.
  • Resultant configuration values are available at the same repository location.
Loading 2103.12269v1…