Research Analyzer
← Back ICRA 2026

A Hybrid Optimization Framework for Grasp Synthesis under Partial Observations

Wenzheng Zhang, Fahira Afzal Maken, Tin Lai, Fabio Ramos

PDF

AI summary

Key figure (auto-extracted from paper)
Integrating a learned energy-based prior with geometric optimization via SVGD significantly boosts grasp success rates from partial point clouds, outperforming existing baselines.
Grasp synthesis Energy-based models Partial observations Stein variational gradient descent Hybrid optimization Robotics

Problem

Grasp synthesis struggles with unseen objects under partial observations, as traditional optimization methods require complete models and pure learning approaches lack geometric robustness and generalization.

Approach

The method trains an Energy-Based Model to score grasp quality and injects its gradient into a Stein Variational Gradient Descent optimizer, which iteratively refines grasp poses alongside an Iterative Closest Point geometric alignment step.

Key results

  • 60.9% average success rate across 5,360 grasp attempts
  • Outperforms AnyGrasp, GPD, and AS-ICP baselines
  • Group-aware data sampling improves energy landscape accuracy
  • Enhanced grasp repeatability and diversity over existing methods

Why it matters

Enables reliable robotic manipulation in real-world scenarios where complete object scans are unavailable, bridging the gap between data-driven learning and geometric optimization.

Abstract

We propose a hybrid grasp synthesis framework that combines a learning-based Energy- Based Model (EBM) with an analytical Iterative Clos- est Point (ICP) method to generate robust grasps from partially observed point clouds. The learned energy function acts as a prior within a Stein Variational Gradient Descent (SVGD) framework, guiding iter- ative refinement of grasp configurations. Evaluated on 67 objects with 5,360 grasp attempts, our method achieves an average success rate of 60.9%, outper- forming AnyGrasp (31.1%) and Grasp Pose Detection (48.4%) and AS-ICP (56.6%). These results highlight the strong generalization ability of our approach and demonstrate how combining data-driven learning with geometric optimization addresses the limitations of either strategy in isolation.

Index terms

Grasping Deep Learning in Grasping and Manipulation Probabilistic Inference

Related papers