← Back ICRA 2026

Plug-And-Play Shape Matching Module for Zero-Shot Mesh-Free Grasp Refinement on Unknown Objects

Juyong Hong, Yeong Gwang Son, Seunghwan Um, Hyouk Ryeol Choi

PDF

AI summary

Key figure (auto-extracted from paper)

A training-free, plug-and-play geometric module boosts grasp stability for unknown objects by estimating centroids from single RGB-D images without CAD models or prior training.

Zero-shot grasping mesh-free refinement RGB-D perception geometric primitive fitting plug-and-play module grasp stability

Problem

Existing grasp planners for logistics automation rely heavily on pre-defined 3D CAD models or large-scale training datasets, making them impractical for novel, unseen objects in real-world environments.

Approach

The system uses an initial grasp candidate to automatically segment the target object from a single RGB-D image, fits geometric primitives to estimate its centroid, and rescoring grasp candidates based on proximity to that centroid to minimize gravitational torque.

Key results

Up to 25% improvement in baseline grasp success rates
Accurate 6D pose and shape estimation without 3D models
Real-time processing latencies of ~165 ms and ~570 ms for different object types
Robust, category-agnostic segmentation and refinement in cluttered scenes

Why it matters

Enables immediate deployment of stable, model-free grasping in logistics and robotics without costly data collection or offline training.

Abstract

Reliably grasping unknown objects in logistics au- tomation remains a major challenge. While most approaches rely on 3D CAD models or large-scale training, their applicability to novel items is limited. This letter proposes a plug-and-play geo- metric refinement module that can be appended to any existing grasp planner. The module operates in a training-free and mesh- free manner, estimating an object’s approximate centroid from a single RGB-D image to enhance grasp stability. Its core mechanism involves using an initial grasp candidate as an automatic prompt for segmentation, followed by geometric primitive fitting to the isolated object’s point cloud. By rescoring grasp candidates based on proximity to the estimated centroid, our module improves phys- ical stability. Experimental results demonstrate that our module improves the success rate of baseline grasp planners by up to 25%p enhancing real-world pick-and-place performance without requiring any offline training or prior object models.

Index terms

Perception for Grasping and Manipulation RGB-D Perception Object Detection Segmentation and Categorization