SoGraB: A Visual Method for Soft Grasping Benchmarking and Evaluation

David Howard, Benjamin G. Greenland, Joshua Pinskier, Xing Wang, Daniel Nguyen, Ge Shi, Tirthankar Bandyopadhyay, Jen Jen Chung, David Howard 0001Robosoft 2025 · DOI

TL;DR

This paper introduces SoGraB, a novel visual benchmarking method for evaluating soft robotic grippers. By quantifying object deformation using 3D point clouds, SoGraB provides a more comprehensive assessment of grasp quality than existing methods, enabling better comparison and design of soft grippers for delicate tasks.

What Problem Does This Paper Address?

The rapid advancement of soft robotic grippers has led to their increasing adoption for handling delicate and deformable objects. However, the field currently lacks a standardized, objective method for evaluating and comparing the performance of different soft gripper designs. Existing evaluation metrics often focus solely on grasp success rate or retention force, neglecting crucial aspects like the forces applied to the object and the resulting deformation or potential for damage. This absence of a common benchmark hinders progress in designing more effective and reliable soft grippers.

Methods

The core of the SoGraB method is its novel approach to quantifying grasp quality by incorporating object deformation. This is achieved through a non-contact, visual technique:

  • 3D Point Cloud Comparison: The system captures 3D point clouds of a soft object before and during grasping.
  • Density-Aware Chamfer Distance (DCD): The deformation is quantified by calculating the DCD between these two point clouds. DCD is chosen over simpler metrics like Chamfer Distance or Earth Mover's Distance because it is robust to variations in point cloud density and the presence of outliers, making it suitable for analyzing potentially incomplete or noisy scans.
  • Scoring Metric: A composite score is derived, combining grasp success (full success, partial success, or failure) with the calculated DCD. For partially successful grasps, the score is further modulated by the proportion of time the object was held before dropping. This results in a scalar benchmark score ranging from 0 (unsuccessful) to 1 (successful).
  • Point Cloud Alignment: To accurately measure deformation, an iterative closest point (ICP) algorithm is used to align the pre- and post-grasp point clouds, correcting for any minor slippage or rotation.

Datasets

The paper introduces a new baseline dataset for soft grasping evaluation:

  • SoGraB Dataset: This dataset comprises 900 grasps, covering 15 object geometries, 3 object materials (Shore 40A, 60A, 85A), and 4 different gripper designs (including three Fin-Ray soft grippers with varying stiffness and one rigid gripper for comparison). This dataset is intended to be expandable by future users.

Key Insights and Ideas

  • Deformation as a Stress Proxy: The paper highlights that object deformation, measured visually, serves as a valuable non-contact proxy for internal stresses and potential damage, offering a more nuanced view of grasp quality than force alone.
  • The "Sweet Spot" for Soft Gripping: SoGraB reveals that soft grippers are most beneficial when their effective stiffness is comparable to that of the object. When objects are too stiff, soft grippers offer little advantage over rigid ones. Conversely, for extremely soft objects, deformation is so significant that gripper stiffness becomes less critical.
  • Object Properties Matter: The stiffness and geometry of the object play a crucial role in determining grasp quality and the relative performance of different grippers. Objects with thin features or complex geometries can lead to higher variability in grasp scores.
  • Generalizability: The visual, non-contact nature of SoGraB makes it applicable to a wide range of gripper-object pairings, provided the object is at least partially visible.

Key Results

  • The SoGraB protocol successfully ranked Fin-Ray gripper designs based on object deformation, validating its ability to differentiate gripper performance.
  • The experimental results showed that grasp scores for soft grippers were generally higher on softer objects, and there was a defined range of object stiffness where soft grippers excelled.
  • The lowest recorded grasp score was 0.517 (Object 03 Shore 40A with Rigid Grippers), and the highest was 0.940 (Object D1 Shore 85A with Rigid Grippers), indicating a broad range of performance outcomes.
  • The study demonstrated that for very stiff objects, the performance difference between soft and rigid grippers was minimal.

Limitations and Future Work

  • Object Visibility: The method relies on the object being at least partially visible, which might limit its application in scenarios with complete occlusion.
  • Dataset Expansion: The authors encourage future work to expand the dataset by including a wider variety of object shapes, materials, and new soft gripper designs.
  • Object Selection Recommendation: For repeatable data, the authors recommend using symmetric objects without thin external features.
  • Future Applications: The authors envision SoGraB as a valuable tool for improving soft gripper design, identifying optimal use cases, and facilitating comparisons with existing and future designs.

Summary updated

BibTeX
@inproceedings{howard2025sograb,
  author = {David Howard and Benjamin G. Greenland and Joshua Pinskier and Xing Wang and Daniel Nguyen and Ge Shi and Tirthankar Bandyopadhyay and Jen Jen Chung and David Howard},
  title = {{SoGraB: A Visual Method for Soft Grasping Benchmarking and Evaluation}},
  booktitle = {Robosoft},
  year = {2025},
  doi = {10.1109/robosoft63089.2025.11020918},
}
Figure